Download layered approach for runtime fault recovery in noc-based

Document related concepts

Computer network wikipedia , lookup

Distributed operating system wikipedia , lookup

RapidIO wikipedia , lookup

Airborne Networking wikipedia , lookup

IEEE 1355 wikipedia , lookup

IEEE 802.1aq wikipedia , lookup

Deep packet inspection wikipedia , lookup

Internet protocol suite wikipedia , lookup

Cracking of wireless networks wikipedia , lookup

Wake-on-LAN wikipedia , lookup

Multiprotocol Label Switching wikipedia , lookup

Recursive InterNetwork Architecture (RINA) wikipedia , lookup

Routing wikipedia , lookup

Routing in delay-tolerant networking wikipedia , lookup

Transcript
PONTIFÍCIA UNIVERSIDADE CATÓLICA DO RIO GRANDE DO SUL
FACULDADE DE INFORMÁTICA
PROGRAMA DE PÓS-GRADUAÇÃO EM CIÊNCIA DA COMPUTAÇÃO
LAYERED APPROACH FOR RUNTIME
FAULT RECOVERY
IN NOC-BASED MPSOCS
EDUARDO WEBER WÄCHTER
Tese apresentada como requisito
parcial à obtenção do grau de Doutor
em Ciência da Computação na
Pontifícia Universidade Católica do Rio
Grande do Sul.
Advisor: Prof. Dr. Fernando Gehm Moraes
Co-Advisor: Prof. Dr. Alexandre de Morais Amory
Porto Alegre, Brasil
2015
Dados Internacionais de Catalogação na Publicação (CIP)
W114l
Wächter, Eduardo Weber
Layered approach for runtime fault recovery in NOC-Based
MPSOCS / Eduardo Weber Wächter. – Porto Alegre, 2015.
91 p.
Tese (Doutorado) – Fac. de Informática, PUCRS.
Orientador: Prof. Dr. Fernando Gehm Moraes
Co-orientador: Prof. Dr. Alexandre de Morais Amory
1. Informática. 2. Arquitetura de Computador.
3. Microprocessadores. I. Moraes, Fernando Gehm. II. Amory,
Alexandre de Morais. III. Título.
CDD 004.35
Ficha Catalográfica elaborada pelo
Setor de Tratamento da Informação da BC-PUCRS
AGRADECIMENTOS
O doutorado é um caminho, longo diga-se de passagem, que não é feito sozinho. São
quatro longos anos de muito trabalho, preocupações e cabelos brancos. Dessa forma
primeiramente agradecer aos meus pais, meus maiores incentivadores. Sei que
independentemente do que aconteça, é pra lá que a gente vai quando a coisa aperta. Eles me
apoiaram muito neste caminho. Amo vocês.
Sempre digo que sem meu orientador essa tese não teria saído. Meu enésimo muito
obrigado ao meu orientador Fernando Moraes, exemplo de pesquisador e professor. Obrigado
pelo tempo que dedicou na tese, nas reuniões, nos brainstorms.
Ao meu co-orientador que conheci no final do mestrado, Alexandre Amory, por trazer mais
“brasa ao assado”. Essa tese tem muita contribuição tua. Muito Obrigado.
Ao apoio financeiro da CAPES/Fapergs, que além da bolsa me proporcionaram uma
experiência única no exterior. Essa experiência me mostrou quais são as prioridades na vida.
Ao Nicolas Ventroux, meu orientador na França e ao chefe do LCE Raphäel David por me
receberem muito bem e me aceitarem no grupo. Serei eternamente grato.
Aos amigos que fiz e que mantenho no laboratório, obrigado pelas boas gargalhadas e
cervejas! Valeu Mandelli, Matheus, Castilhos, Guindani, Raupp, Julian, Ost, Leonel, Edson,
Leonardo, Augusto, Carlos Henrique, Walter, Ruaro, Felipe, Heck, Sérgio, Thiago, Madalozzo.
Aos integrantes e amigos do FBC e da Xorna, o meu muito obrigado por me tirar da rotina
quando eu precisava, e algumas vezes quando eu não precisava também!
Por fim, gostaria de agradecer ao amor da minha vida. A pessoa que faz os meus dias
serem mais felizes, Keli, te amo. Obrigado por dizer que tudo ia dar certo, quando não sabia se ia
dar certo. Obrigado por acreditar em mim.
ABORDAGEM EM CAMADAS PARA RECUPERAÇÃO DE FALHAS EM
TEMPO DE EXECUÇÃO EM MPSOCS BASEADOS EM NOC
RESUMO
Mecanismos de tolerância a falhas em MPSoCs são obrigatórios para enfrentar defeitos
ocorridos durante a fabricação ou falhas durante a vida útil do circuito integrado. Por exemplo,
falhas permanentes na rede de interconexão do MPSoC podem interromper aplicações mesmo
que a rede tenha caminhos sem falha para um determinado destino. A tolerância a falhas em
tempo de execução fornece mecanismos de auto-organização para continuar a oferecer serviços
de processamento apesar de núcleos defeituosos devido à presença de falhas permanentes e/ou
transitórias durante toda a vida dos chips. Esta Tese apresenta uma abordagem em camadas
para um MPSoC tolerante a falhas, onde cada camada é responsável por resolver uma parte do
problema. O método é construído sobre uma nova proposta de rede especializada utilizada para
procurar caminhos livre de falha. A primeira camada, denominada camada física, é responsável
pela detecção de falhas e isolamento das partes defeituosas da rede. A segunda camada,
denominada camada de rede, é responsável por substituir um caminho defeituoso por um
caminho alternativo livre de falhas. Um método de roteamento tolerante a falhas executa o
mecanismo de busca de caminhos e reconfigura a rede para usar este caminho livre de falhas. A
terceira camada, denominada camada de transporte, implementa um protocolo de comunicação
tolerante a falhas que detecta quando pacotes não são entregues ao destino, acionando o método
proposto na camada de rede. A última camada, camada de aplicação, é responsável por mover as
tarefas do elemento de processamento (PE) defeituoso para um PE saudável, salvar o estado
interno da tarefa, e restaurá-la em caso de falha durante a execução. Os resultados na camada de
rede mostram um método rápido para encontrar caminhos livres de falhas. O processo de procura
de caminhos alternativos leva tipicamente menos de 2000 ciclos de relógio (ou 20
microssegundos). Na camada de transporte, diferentes abordagens foram avaliadas para detectar
uma mensagem não entregue e acionar a retransmissão. Os resultados mostram que a
sobrecarga para retransmitir a mensagem é 2,46 vezes maior quando comparado com o tempo
para transmitir uma mensagem sem falha, sendo que todas outras mensagens subsequentes são
transmitidas sem sobrecarga. Para as aplicações DTW, MPEG e sintética, o caso médio de
sobrecarga no tempo de execução da aplicação é de 0,17%, 0,09% e 0,42%, respectivamente.
Isto representa menos do que 5% do tempo de execução de uma dada aplicação no pior caso. Na
camada de aplicação, todo o protocolo de recuperação de falhas executa rapidamente, com uma
baixa sobrecarga no tempo de execução sem falhas (5,67%) e com falhas (17,33% - 28,34%).
Palavras Chave: Sistemas Multi-Processados em Chip (MPSoC), Redes Intra-chip (NoC),
Tolerância a Falhas.
LAYERED APPROACH FOR RUNTIME FAULT RECOVERY
IN NOC-BASED MPSOCS
ABSTRACT
Mechanisms for fault-tolerance in MPSoCs are mandatory to cope with defects during
fabrication or faults during product lifetime. For instance, permanent faults on the interconnect
network can stall or crash applications, even though the MPSoCs’ network has alternative faultfree paths to a given destination. Runtime Fault Tolerance provide self-organization mechanisms
to continue delivering their processing services despite defective cores due to the presence of
permanent and/or transient faults throughout their lifetime. This Thesis presents a runtime layered
approach to a fault-tolerant MPSoC, where each layer is responsible for solving one part of the
problem. The approach is built on top of a novel small specialized network used to search fault-free
paths. The first layer, named physical layer, is responsible for the fault detection and fault isolation
of defective routers. The second layer, named the network layer, is responsible for replacing the
original faulty path by an alternative fault-free path. A fault-tolerant routing method executes a path
search mechanism and reconfigures the network to use the faulty-free path. The third layer, named
transport layer, implements a fault-tolerant communication protocol that triggers the path search in
the network layer when a packet does not reach its destination. The last layer, application layer, is
responsible for moving tasks from the defective processing element (PE) to a healthy PE, saving
the task’s internal state, and restoring it in case of fault while executing a task. Results at the
network layer, show a fast path finding method. The entire process of finding alternative paths
takes typically less than 2000 clock cycles or 20 microseconds. In the transport layer, different
approaches were evaluated being capable of detecting a lost message and start the
retransmission. The results show that the overhead to retransmit the message is 2.46X compared
to the time to transmit a message without fault, being all other messages transmitted with no
overhead. For the DTW, MPEG, and synthetic applications the average-case application execution
overhead was 0.17%, 0.09%, and 0.42%, respectively. This represents less than 5% of the
application execution overhead worst case. At the application layer, the entire fault recovery
protocol executes fast, with a low execution time overhead with no faults (5.67%) and with faults
(17.33% - 28.34%).
Keywords: Multiprocessor System-on-Chip (MPSoC), Network-on-Chip (NoC), Fault Tolerance.
LIST OF FIGURES
Figure1–Evolutionoftransistornumber,clockfrequency,powerconsumptionandperformance.Adapted
from[SUT13]...................................................................................................................................................17
Figure2–Failureprobabilitytrendsasafunctionofthetechnologynode[SAN10].....................................19
Figure3–Motivationalexampletojustifytheadoptionofalayeredapproachforafault-tolerantMPSoC.
RedPEsrepresentthepathtakenbypacketsassuminganXYroutingalgorithm.GreenPEsrepresentsa
possiblenewpathtothedestination.ThenumbersdenotetheThesis’chapterdetailingeachlayer..........20
Figure4–MPSoC,PE,andRoutermodels.TheManagerProcessorislabeledasMPandtheSlaveProcessor
asSP................................................................................................................................................................23
Figure5–DetailedviewoftheNoCwithduplicatedphysicalchannelandinputbuffers[CAR09]................24
Figure6-Applicationmodeledasataskgraph..............................................................................................24
Figure7–CommunicationprotocoladoptedinthereferenceMPSoC..........................................................25
Figure8–BasicperformanceconceptsindistributedsystemsthatwillbeusedinthisThesis.....................26
Figure9–Fault-TolerantRouterarchitectureproposal..................................................................................28
Figure10–Portswappingmechanism[FIC09]................................................................................................29
Figure11–In(a)thearchitectureofthetestwrappercell(TW),withanisolationmechanismcontrolledby
thefaultsignal.In(b)thelocationofeachTWattherouter.........................................................................30
Figure12–TestWrapperscontrolledbyafaultsignal.dfa.............................................................................31
Figure13–PEarchitecture,detailingtheinternalrouterstructure(TW:testwrappers)..............................31
Figure14–InterPEarchitecture,detailingtheinterconnectionsandfaultlocationsintherouter..............32
Figure15-Exampleofregiondefinitions[FLI07]............................................................................................34
Figure16-Thepathfollowedbyapacketaroundthefault(blueandredlines).Thewhitenodeisthe
exclusionregion,thegraynodesarethereconfiguredasboundaryrouters.................................................36
Figure17–DMeshtopology...........................................................................................................................36
Figure18–Connectionmodificationtomaintainconnectivityinpresenceoffaults.....................................37
Figure19–MiCoFRoutinginscenarioswithtwofaultyrouters[EBR13].......................................................37
Figure20–Regioncomputationprocessproposedby[FLI07].......................................................................39
Figure21–Flowchartoftheroutingmethod.................................................................................................41
Figure22–Seekstepsofthemethodina4x4mesh,withfourfaultyrouters...............................................42
Figure23–Backtrackingprocessoftheproposedmethod.............................................................................42
Figure24–Inter-routerconnectionoftheSeekrouter..................................................................................43
Figure25–Intra-routerconnectionoftheSeekrouter..................................................................................44
Figure26–SpidergonSTNoC,topologyandphysicalview.............................................................................45
Figure27–Hierarchical-Spidergontopology.Redlinesshowtheconnectionbetweenthelowerandupper
layers...............................................................................................................................................................45
Figure28–Configurationusedinmeshandtorusnetworks.Grayroutersarefaultyroutersandthedark
grayroutersarecommunicatingrouterpairs.Fourcommunicatingpairsareillustrated:81à9,45à47,
41à49,9à55.................................................................................................................................................45
Figure29–Minimalpathfoundbytheproposedmethodfromrouter81to09(darkgray),inatorus
topology(grayroutersarefaultynode).Thearrowsindicatewraparoundlinks...........................................46
Figure30–SpidergonTopology(grayroutersarefaultynodes).Diagonalconnectionswereomitted.Inthe
detailwehavethefourpairsofevaluatedflows............................................................................................46
Figure31–EvaluationscenarioofaNoC10x10.............................................................................................48
Figure32–TheproposedcommunicationprotocoltoprovideFT.................................................................54
Figure33–Protocoldiagramofamessagedeliveryfault..............................................................................54
Figure34–Protocoldiagramofackmessagefault........................................................................................55
Figure35–Protocoldiagramofmessagerequestfault.................................................................................55
Figure36–Exampleofafaultyrouterinthemiddleofamessagebeingdelivered......................................56
Figure37–Comparingfaultyandfault-freeprotocollatency........................................................................58
Figure38–Auto-detectionoffaultypathsprocess........................................................................................59
Figure39–SequencechartofprotocolcommunicationwithAuto-detectionoffaultypaths.......................60
Figure40–Exampleofacknowledgmentmessagefollowedbyamessagerequest......................................60
Figure41–FTcommunicationprotocolwithouttheacknowledgmentmessage..........................................61
Figure42–Faultsimulationflow....................................................................................................................62
Figure43–Evaluatedapplications,theirtaskmappingsandtaskgraphs......................................................63
Figure44–Taskisolationduetofaultsateastandsouthports.....................................................................64
Figure45–Applicationexecutiontime(AET)forbasicapplicationinscenarioswithfaultsasafunctionof
parameterk.Eachlinerepresentsagivenscenariowithonefault................................................................65
Figure46–Applicationexecutiontime(AET)forMPEGapplicationinscenarioswithfaultsasafunctionof
parameterk.Eachlinerepresentsagivenscenariowithonefault................................................................65
Figure47–Applicationexecutiontime(AET)forsyntheticapplicationinscenarioswithfaultsasafunction
ofparameterk.Eachlinerepresentsagivenscenariowithonefault............................................................66
Figure48–Timespenttotransmiteight256-flitpacketswithandwithoutfaultusingtheproposedFT
communicationprotocolforthe8firstframes.Thefaultwasdetectedinthethirdpacket.........................67
Figure49–Applicationsexecutiontimeforeachfaultyscenario.Eachdotinthefigurerepresentsagiven
scenario...........................................................................................................................................................68
Figure50–P2012architecture.......................................................................................................................75
Figure51–ExecutionModelinP2012withHARS.(1)masterforksparalleltasks,(2)otherPEsexecutethe
tasks,and(3)themasterdoesthejoin...........................................................................................................76
Figure52–Fault-TolerantExecutionModel:In(1)themasterexecutesacontextsavingandin(3)itverifyif
therewasafault,ifpositive,thecontextisrestoredandtheforkisre-executedavoidingthefaultyPE......78
Figure53–TaskGraphofthesyntheticapplication.(1)ThePEmexecutesthecontextsaving,(2)thefork
splitstheexecutioninTtasks,eachoneexecutingNnumberofNOPinstructions,and(3)thisprocessis
replicatedRtimes...........................................................................................................................................78
Figure54–ExecutiontimeoverheadvaryingthenumberofNOPsineachtask(T=10,R=10)......................79
Figure55–Executiontimeoverheadofcontextsavingchangingthecontextdatasizefrom10to10kwords
of32bits(T=10,R=10,N=10,000)...................................................................................................................79
Figure56–Forkoverheadvaryingthenumberofrepetitions(T=10,N=10,000)...........................................80
Figure57–Applicationexecutiontimeoverheadforscenarioswithnocontextsaving,andtheoverheadfor
scenarioswherethereisoverheadincreasingthenumberoffaults(T=10,R=10,N=10,000).......................80
Figure58–Executiontimeoverheadwithoutfaultswhenexecutingcontextsavingfromeachframetoeach
16frames.Thebarsshowthecontextsavingoverheadandtheexecutiontime..........................................81
Figure59–Applicationexecutiontimewithnocontextsaving,theoverheadinducedbythecontextsaving
andtheoverheadinducedbythecontextsavingplustherecoverytimeforone,twoandthreefaults.The
percentagesrepresenttheoverheadcomparedtothebaseline.Thehighlightedpartrepresentsthetime
executedrestoringthecontext.......................................................................................................................81
LIST OF TABLES
Table1–EvolutionofnumberofPEsandinterconnection............................................................................18
Table2–Comparisonofdifferentroutingapproaches..................................................................................39
Table3–Timetoexecutetheproposedroutingmethod,inclockcycles,fordifferenttopologies...............47
Table4–AreaoverheadforXilinxxc5vlx330tff1738-2device.......................................................................49
Table5–Comparisonofdifferentcommunicationprotocolsadoptingmessagepassing.............................53
Table6–Validationresultswithonefault......................................................................................................63
Table7–Validationresultswithtwofaults....................................................................................................63
Table8–NumberofclockcyclesforeachstepoftheFTcommunicationprotocolofFigure39varyingthe
numberofhopsbetweenPE1andPE2............................................................................................................67
Table9–AETworst-caseoverheadforWatchdogtimersandthetwoAuto-detectionofFaultyPaths
approaches......................................................................................................................................................69
Table10–ComparisonofevaluatedFault-TolerantContextSavingapproaches..........................................74
Table11–AccesstimeformemoriesinP2012..............................................................................................77
Table12–SimplifiedviewoftheOSImodelwiththeaddedfeaturesforfaulttoleranceineachlayer........84
Table13–PublicationsduringthePhDperiod...............................................................................................91
LIST OF ABBREVIATIONS
BIST
CRC
CMOS
DTW
ECC
FF
FIFO
FPGA
FT
GALS
GPP
GPU
HeMPS
ILP
IP
LUT
MIPS
MPEG
MPI
MPSoC
NI
NoC
OS
OSI
PE
SEU
SoC
TFAP
TMR
TCDM
UCD
VHDL
VHSIC
VLSI
Built-in self-test
Cyclic redundancy check
Complementary metal-oxide semiconductor
Dynamic Time Warping
Error Control Coding
Flip-Flop
First-In First-Out
Field Programmable Gate Arrays
Fault-Tolerant
Globally Asynchronous Locally Synchronous
General Purpose Processor
Graphics Processing Unit
Hermes Multiprocessor System
Instruction-Level Parallelism
Intellectual Property
Look up Table
Microprocessor without Interlocked Pipeline Stages
Moving Picture Experts Group
Message Passing Interface
Multi-Processor System-on-Chip
Network Interface
Network-on-Chip
Operating System
Open Systems Interconnection model
Processing Element
Single Event Upset
System-on-Chip
Time to Find Alternative Paths
Triple Module Redundancy
Tightly Coupled Data Memory
Unresponsive Communication Detection
VHSIC Hardware Description Language
Very High Speed Integrated Circuit
Very-Large-Scale integration
SUMMARY
1
INTRODUCTION.......................................................................................................................................17
1.1
OverviewoftheThesis....................................................................................................................20
1.2
Goals................................................................................................................................................21
1.3
OriginalityofthisThesis..................................................................................................................21
1.4
HypothesistobedemonstratedwiththisThesis............................................................................21
1.5
StructureofthisDocument.............................................................................................................22
2 REFERENCEARCHITECTUREANDBASICCONCEPTS................................................................................23
2.1
ReferenceArchitecture...................................................................................................................23
2.1.1
TaskCommunication...............................................................................................................24
2.2
BasicFaultToleranceConcepts.......................................................................................................26
2.3
Generalassumptionsrelatedtothearchitectureandfaultmodel.................................................27
3 PHYSICALLAYER......................................................................................................................................28
3.1
ExamplesofPhysicalLayerApproaches..........................................................................................28
3.2
PhysicalLayerProposal...................................................................................................................29
3.3
FinalRemarks..................................................................................................................................32
4 FAULT-TOLERANTROUTINGMETHOD–NETWORKLAYER.....................................................................33
4.1
State-of-the-ArtinFault-TolerantRoutingAlgorithmsforNoCs.....................................................33
4.1.1
FDWR[SCH07].........................................................................................................................33
4.1.2
Flichetal.[FLI07].....................................................................................................................34
4.1.3
FTDR[FEN10]...........................................................................................................................35
4.1.4
uLBDR[ROD11]........................................................................................................................35
4.1.5
iFDOR[SEM11]........................................................................................................................35
4.1.6
Vicis[DEO12]...........................................................................................................................36
4.1.7
Alhussienetal.[ALH13]...........................................................................................................36
4.1.8
MiCoF[EBR13].........................................................................................................................37
4.1.9
BLINC[DOO14]........................................................................................................................38
4.2
Discussiononthestate-of-the-art...................................................................................................38
4.3
ProposedApproach:MAZENOCRoutingMethod[WAC12b][WAC13a].........................................40
4.3.1
HardwareStructure.................................................................................................................43
4.4
Results.............................................................................................................................................44
4.4.1
ExperimentalSetup.................................................................................................................44
4.4.2
TimetoFindAlternativePaths(TFAP).....................................................................................47
4.4.3
MultipleSeeksEvaluation.......................................................................................................48
4.4.4
AreaOverhead........................................................................................................................48
4.5
FinalRemarks..................................................................................................................................49
5 FAULT-TOLERANTCOMMUNICATIONPROTOCOL–TRANSPORTLAYER................................................51
5.1
TheState-of-the-ArtinProtocolandSystemReconfigurationApproaches....................................51
5.2
ProposedFault-TolerantCommunicationProtocolDescription.....................................................53
5.3
FaultsintheProposedFTCommunicationProtocol.......................................................................54
5.3.1
FaultintheMessageDelivery.................................................................................................54
5.3.2
FaultinMessageAcknowledgment.........................................................................................54
5.3.3
FaultinMessageRequest........................................................................................................55
5.3.4
FaultWhileReceivingPackets.................................................................................................56
5.4
Fault-TolerantCommunicationProtocolImplementation..............................................................56
5.4.1
AdaptiveWatchdogTimers[WAC14a]....................................................................................57
5.4.2
Auto-detectionoffaultypaths[WAC14b]...............................................................................58
5.4.3
Auto-detectionoffaultypathswithsimplifiedprotocol.........................................................60
5.5
Results.............................................................................................................................................61
5.5.1
PerformanceEvaluationofAdaptiveWatchdogTimersImplementation[WAC14a].............64
5.5.2
PerformanceEvaluationofAuto-detectionoffaultypaths[WAC14b]...................................66
5.5.3
PerformanceEvaluationofAuto-detectionoffaultypathswithsimplifiedprotocol.............69
5.5.4
OverallEvaluationoftheFTCommunicationProtocols..........................................................69
5.6
FinalRemarks..................................................................................................................................69
6 APPLICATIONRECOVERY.........................................................................................................................71
6.1
TheState-of-the-ArtinContextSavingandRecoveryTechniques.................................................71
6.1.1
ReVive[PRV02]........................................................................................................................71
6.1.2
Chip-levelRedundantThreading[GON08]..............................................................................72
6.1.3
Reli[LI12].................................................................................................................................72
6.1.4
DeSyRe[SOU13]......................................................................................................................72
6.1.5
Barreto’sApproach[BAR15]....................................................................................................73
6.1.6
Rusu’sApproach[RUS08]........................................................................................................73
6.1.7
Stateoftheartdiscussion.......................................................................................................74
6.2
Fault-tolerantReferencePlatform..................................................................................................74
6.2.1
Architecture.............................................................................................................................75
6.2.2
SoftwareStack.........................................................................................................................75
6.2.3
Executionmodel......................................................................................................................76
6.3
ProposedFault-TolerantMethod[WAC15]....................................................................................77
6.4
EvaluationoftheProposedFault-TolerantMethod.......................................................................78
6.4.1
EvaluationofthemethodwithSyntheticApplication.............................................................79
6.4.2
EvaluationofthemethodwithHBDCApplication..................................................................80
6.5
FinalRemarks..................................................................................................................................82
CONCLUSION...........................................................................................................................................83
7.1
LimitationsoftheCurrentProposal................................................................................................84
7.2
FutureWorks...................................................................................................................................84
REFERENCES....................................................................................................................................................86
APPENDIX1–LISTOFPUBLICATIONS.............................................................................................................91
7
17
1 INTRODUCTION
From many years to now, processors evolved from single cores with thousands of
transistors to many cores with billions of these devices. This evolution enabled the aggressive
growth in the number of transistors and the integration of complete systems into a single die.
These systems, named Systems-on-Chip (SoCs), are widely used in electronic devices, as in
multimedia and telecommunication applications, where high performance and power consumption
are important design constraints.
In the past decades, the increase of performance was simply achieved by increasing the
frequency of the SoC. However, today a single processor system may not provide the required
performance for embedded applications [JER05]. Several “walls” explain this limitation:
•
ILP (Instruction Level Parallelism) wall, even with aggressive speculative techniques
and large superscalar datapaths, the number of simultaneous instructions executed per
clock cycle reached a limit;
•
Frequency and power walls, the power increases linearly with the frequency. Power
dissipation limits the SoC frequency. Even if CMOS technologies enable operating
frequencies superior to 5 GHz, SoCs work at much lower frequencies due to power
dissipation issues.
As Figure 1 shows, even if the number of transistors integrated on a single die grows
according to Moore’s law, the frequency and power do not follow this trend.
Figure 1 – Evolution of transistor number, clock frequency, power consumption and performance.
Adapted from [SUT13].
The alternative to these design and performance gaps is to include more processors in the
same die. As a result, instead of using a single processor with a high frequency, new architectures
with many simpler and slower processors are used to fill those gaps. These architectures are
named Multiprocessor SoCs (MPSoCs). The MPSoC approach raises the performance using a
coarse grain parallelism, extending the quest for higher ILP by parallel tasks executing on multiple
processors.
18
The MPSoC trend is nowadays well established, being adopted in GPP (general-purpose
processors) and GPU (graphics processing unit) markets. There are many examples of such
products: PlayStation 3, which uses the Cell processor [IBM13], with 9 PEs (Processing Elements);
Oracle Sparc T5 [FEE13], with 16 PEs. Other examples of MPSoCs include TilePro64 [TIL13] (up
to 72 PEs), CSX700 [CLE13] (192 PEs), Teraflops [INT13a] (80 PEs), NVIDIA’s Kepler TM GK110
[NVI13] (2,880 PEs) and Single-Chip Cloud Computer [INT13b] (48 PEs). The perspective is that
hundreds of PEs in a single die will become mainstream at the end of the decade [ITR15].
As the number of PEs increases, the challenge is to interconnect such PEs in an efficient
and scalable fashion. As can be observed in Table 1, the interconnection of large MPSoCs shifts
to NoCs (networks-on-chip) [BOR07]. On the other hand, most of the GPUs are an STMD (Single
Thread Multiple Data) architecture where all threads execute the same code. For this reason,
general-purpose systems adopt message passing (NoCs) and specialized systems, as GPUs,
adopt shared memory (hierarchical busses).
The interconnection architectures have evolved from point-to-point links to single and
multiple hierarchical buses, and to Networks-on-Chip (NoC) [BEN02]. NoCs are composed of
cores connected to routers, and routers interconnected by communication channels [BEN02]. The
use of NoCs as the communication infrastructure in these platforms is a reality in academic
research and industrial projects, considered an alternative to traditional buses, offering as major
advantages scalability and support to multiple parallel communications. However, the motivation
for their use in SoCs goes beyond these advantages. NoCs can support many communication
services with different levels of quality of service, and offer intrinsically fault tolerance support
(more than one path for a source-target pair). The present work assumes general-purpose
systems, and hence, NoCs are adopted as the communication infrastructure.
Table 1 – Evolution of number of PEs and interconnection.
Chip
# of PEs
Year
Interconnection
Cell
9
2006
Ring
Teraflops
80
2007
NoC
TilePro64
64
2008
NoC
Single-Chip Cloud Computer
48
2009
NoC
KeplerTM GK110
2880
2012
Hierarchical bus
Oracle Sparc T5
16
2013
Crossbar
CSX700
192
2013
NoC
It is important to differentiate HPC (High-Performance Computing) from embedded systems.
In HPC, performance is the main cost function, not power dissipation and energy consumption. An
off-chip network interconnects thousands of high-end processors, providing high-performance
results. The problem is that, depending on the number of processors, you might need an electrical
substation just to deliver the power required by the high-end processors. On the other hand,
embedded systems have different constraints, as reduced power consumption, small size, and low
cost-per-unit. The present work assumes the design of embedded systems, where such
constraints must be fulfilled, together with the performance required by the workload of current
embedded applications, like multimedia, telecommunication protocols (as LTE), GPS, augmented
reality [ASA06].
19
In one hand, a larger number of devices in a single die enable the design of MPSoCs, as
previously mentioned. On the other hand, the integration of many PEs in a single die reduces the
reliability of the system since the probability of faults is related to the transistor channel length and
the area of the integrated circuit for a given technology node. In addition, the failure probability is
getting higher due to technology miniaturization (Figure 2). Note that same color lines represent
failure probability for 1, 5 and 10 years estimation. The combination of the following factors
increases the failure rate: (i) higher transistors integration; (ii) lower voltage levels; (iii) higher
operation frequency. In highly scaled technology, a variety of wear-out mechanisms can cause
transistor failures. As transistor dimensions are getting smaller, effects like oxide breakdown
become a concern since the gate oxide tends to become less effective over time. Moreover,
negative bias temperature instability [ALA03] is of special concern in PMOS devices, where there
is increased threshold voltage over the time. Additionally, thin wires are susceptible to
electromigration [GHA82] because conductor material is gradually worn away during chip operation
until an open circuit occurs. Since these mechanisms occur over the time, traditional burn-in
procedures and manufacturing tests are ineffective to detect them.
Figure 2 – Failure probability trends as a function of the technology node [SAN10].
All these factors threaten the reliability of future computational systems. Therefore, there is
a need to adopt fault-tolerant techniques in the MPSoC design, to maintain the system operation
even in the presence of faults. Fault Tolerance (FT) can be defined as the property that enables a
system to continue to operate correctly in the presence of a failure in some of its components.
Section 1.1 describes an example of how a single fault in a single processor of a MPSoC can
compromise the whole system.
NoC-based MPSoCs provide excellent opportunities for the design of reliable systems. For
example, there are redundant paths between processors and it is possible to design an integrated
hardware-software fault-tolerant approach, combining the hardware performance with the software
flexibility. Furthermore, according to [HEN13], “Runtime adaption will be another key component to
increase the reliability of future on-chip systems”. If we employ only design-time techniques,
designs would have to be more and more conservative as they would have to reserve more and
more resources to provide some infrastructure to support failures of components. Consequently,
applying only design-time techniques would render future on-chip systems infeasible from the cost
point of view, since the failure rate will continue to increase with future technology nodes.
In this context, this Thesis proposes a layered approach for runtime fault recovery for NoCbased MPSoCs. The method is built on top of a small specialized network used to search fault-free
paths, and an MPI-like software protocol, implemented at the transport layer, hidden from the
application layer. The software protocol detects unresponsive processors and automatically fires
20
the path search network, which can quickly return a new path to the target processor. This
approach has a distributed nature since a source processor uses only local information to monitor
the path to the targets. Thus, there is no single point of failure in the system and the approach is
scalable in terms of number of PEs. The proposed approach provides complete reachability, i.e. if
there is only one fault-free path to the target, it will be found, even in the presence of multiple
faults. At the application level, this Thesis presents a checkpoint and rollback mechanism enabling
the recovering of an application if a PE fails.
1.1
Overview of the Thesis
This section exposes the need for a layered approach to achieve a FT MPSoC and
summarizes each layer of the approach. For instance, let us assume that a fault occurs in the path
between tasks A and B, mapped at different PEs, as illustrated in Figure 3. Since a conventional
message passing library has no FT mechanism (e.g. timeout), a Receive() call can wait indefinitely
if an unreliable communication layer loses the message. This blockage can increase the
communication buffer’s occupancy, blocking other pairs of communicating tasks, ultimately
resulting in a system crash caused by a single fault. In addition, a faulty packet might be misrouted
or occupy buffers indefinitely, and a faulty router might generate wrong signals to its neighbors
(Byzantine fault) leading to invalid system state.
Transport Layer
- FT Communication Protocol
- Fault Free Path Storage
- Path Search Triggering
Application Layer
- PE disabling
- Save task internal state
(checkpoints)
- Restore task (rollback)
5
6
[WAC14a][WAC14b]
A
PE
Network Layer
- Fault Tolerant Routing
- New Path Search
- Path Reconfiguration
4
[WAC12b][WAC13a]
B
[WAC15]
Physical Layer
- Link Fault Detection
- Fault Isolation (Link
Wrapper)
3
Figure 3 – Motivational example to justify the adoption of a layered approach for a fault-tolerant
MPSoC. Red PEs represent the path taken by packets assuming an XY routing algorithm. Green PEs
represents a possible new path to the destination. The numbers denote the Thesis’ chapter detailing
each layer.
Therefore, to avoid system crashes induced by hardware faults, the system must detect,
isolate, and avoid the faulty region. For instance, if the fault is permanent and the network uses
only deterministic routing algorithm, then the packet retransmission would use the same faulty
path, generating a new undelivered packet. Thus, the system must have the ability to locate the
faulty router and support adaptive routing, delivering packets using an alternative path and
avoiding the faulty region. To solve this problem, we adopted a divide and conquer approach,
where each layer is responsible for solving a given part of the problem.
The proposed layered approach is divided into four main layers. Each layer is numbered
according to one Chapter of this Thesis. The first layer (Figure 3 – 3), named physical layer, is
responsible for the fault detection mechanism and the wrapper module used for fault isolation
purposes. The second layer (Figure 3 – 4), named the network layer, is responsible for replacing
the original faulty path by an alternative healthy path. A fault-tolerant routing method that executes
21
a path search mechanism enables this process [WAC12b][WAC13a]. The third layer (Figure 3 – 5),
named transport layer, implements a fault-tolerant communication protocol that triggers the path
search in the network layer [WAC14a][WAC14b]. The last layer (Figure 3 – 6), application layer is
responsible for moving tasks from the defective PE to a healthy PE, saving the task’s internal state
and restoring it in case of fault while executing a task (i.e. checkpoints and rollback) [WAC15].
1.2
Goals
The strategic goals of this Thesis include:
• Get a holistic understanding of FT NoC-based MPSoC design, from the physical level
to the system level;
•
Explore the use of routing algorithms for fault-tolerant NoCs;
•
Explore the use of methods for isolation of faulty modules;
•
Define a fault-tolerant communication protocol for NoC-based MPSoCs;
•
Explore application recovery methods when faults are detected in PEs;
•
Task remapping in the presence of faults in PEs.
To accomplish these strategic goals, the following specific objectives should be fulfilled:
• Propose a new fault-tolerant routing method for NoCs;
1.3
•
Propose a method to isolate faulty modules;
•
Reconfigure the network using a new faulty-free path;
•
Change the communication protocol, adding fault-tolerant features on it;
•
Addition of a method to detect undelivered messages;
•
Propose a method for task remapping in case of faults in a processor executing a task.
Originality of this Thesis
The originality of this Thesis relies in the layered approach for achieving fault tolerance in
MPSoCs. Each layer is responsible for solving a part of the problem, adopting a divide and
conquer approach. At the network layer, we propose a novel fault-tolerant routing method, being
the first in the state-of-the-art with full reachability and network topology agnostic. At the network
layer, we propose and evaluate three approaches for delivering messages even in the presence of
faults in the NoC. At the application layer, we present a context saving approach to isolate and
restart faulty processors for a state of the art MPSoC.
1.4
Hypothesis to be demonstrated with this Thesis
This Thesis shows that a layered approach can cope with faults in the communication
infrastructure and in the processing elements in such a way to ensure a correct operation of the
MPSoC, even in the presence of multiple faults. The proposed methods increase the lifetime of
current MPSoCs due to the isolation of defective routers, links, and processing elements. The
remaining healthy components continue to execute the applications correctly.
22
1.5
Structure of this Document
This Thesis adopts self-contained Chapters, i.e., each Chapter presents a revision of the
state of the art, the approach proposed by the Author and the results with the corresponding
discussion. This approach is similar to the OSI standard, where each layer implements its functions
and services assuming a lower layer. For example, the communication protocol (Chapter 5)
assumes that there is a lower layer (Chapter 4 – Network layer) with a path search method.
The remaining of this document is organized as follows. Chapter 2 presents basic concepts
and presents the reference platform used to validate the proposed methods. Next, Chapter 3
presents the physical layer, where network faults are detected and isolated. Chapter 4 presents the
network layer, with the proposal of an FT routing method for NoCs. Chapter 5 presents the
transport layer, with proposals for an FT communication protocol. Chapter 6 presents the last FT
layer, the application layer. Finally, Chapter 7 concludes the Thesis bringing a discussion of strong
and weak points of the proposed layered approach and providing some insights for future works.
23
2 REFERENCE ARCHITECTURE AND BASIC CONCEPTS
This Chapter discusses three main topics related to this Thesis:
(i)
The basic features of the reference architecture, such as: tasks communication, task
mapping and task execution;
(ii) Basic concepts of fault tolerance, focusing on fault tolerance at the system architecture
level;
(iii) General assumptions for the fault-tolerant methods proposed in this work.
2.1
Reference Architecture
The reference architecture is the HeMPS MPSoC [WOS07][CAR09], which is an array of
PEs interconnected by a mesh NoC [CAR10] (Figure 4(a)).
Each PE contains an IP connected to a NoC router, as illustrated in Figure 4(b). The IP has
the following modules: (i) a 32-bit processor (MIPS-like architecture); (ii) a dual-port local memory;
(iii) a DMA module, enabling parallelism between computation and communication; (iv) a network
interface. The Manager Processor (MP in the Figure) controls the system and the applications. The
MP also accesses the task repository memory, an external memory that stores the object codes of
the applications. Slave processors (SP) execute user tasks. The hardware of the SPs and MP is
the same, being the differentiation made by the software running at each PE. Each SP may
execute multiple tasks and communicate with neighbors PEs through the NoC.
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
SP
MP
SP
SP
SP
SP
SP
(a) MPSoC model
CPU
Router
Network
Interface
SP
RAM
SP
DMA
(b) PE model
North Port
T
TWW
Local
Port
TW
TW
Task
Repository
PE (SP or MP)
East
port
ch 0
West link 0 ch 1
port link 1 ch 0
ch 1
Crossbar
(10 x 10)
TW
TW
TW
TW
TW
TW
Arbitration
and Routing
South port
(c) Router model
Figure 4 – MPSoC, PE, and Router models. The Manager Processor is labeled as MP and the Slave
Processor as SP.
The Figure 4(c) presents the definitions for interconnections between routers used in this
document. The wires (16 in our case) connecting one router to another are named channel. Each
router contains two output channels and two input channels. The proposed router is composed of
duplicated physical channels, being the two links combined named links. The aggregation of two
links is named port.
The NoC adopts a 2D-mesh topology, with input buffering, credit-based flow control, and
duplicated physical channels. The standard routing mode between PEs is the distributed XY
24
routing algorithm. The network also supports source routing such that it is possible to determine
alternative paths to circumvent hot spots or faulty areas (feature discussed in details in Chapter 4).
The NoC was originally designed for QoS purposes with duplicated physical channels and input
buffers, as presented in Figure 5. As discussed in [CAR08], using duplicated physical channels
compared to virtual channels (2 lanes) saves 12% in the area for a single router. For FT purposes,
it is attractive to use duplicated physical channels. If a fault is detected at a given link (wires, input
buffers or an output port), there is an alternative link capable to preserve the communication
between routers.
Router
Switch Control
Input buffers
Crossbar 10x10
North0
North1
South0
South1
East0
East1
West0
West1
Local0
Local1
Local1
Local0
Wes t1
West0
East1
East0
South1
South0
North1
North0
Figure 5 – Detailed view of the NoC with duplicated physical channel and input buffers [CAR09].
At the software level, each processor executes a small operating system (kernel),
responsible for communication among tasks, management services, and multi-task execution.
Message passing is the communication method adopted since it is more suited for distributed
systems with local memory. Applications are modeled as task graphs A=<T,C> (example in Figure
6), where T = {t1, t2, ..., tm} is the set of application tasks corresponding to the graph vertices, and
C= {(ti, tj, wij) | (ti, tj) ∈ T and wij ∈ ℕ*} denotes the communications between tasks, corresponding to
the graph edges (wij corresponds to the communication volume between ti, and tj). The MPSoC
assumes that there is an external MPSoC memory, named task repository, with all applications
tasks (set T) loaded into the system at runtime.
T2
T4
T1
Initial
task
T5
T3
Figure 6 - Application modeled as a task graph.
2.1.1 Task Communication
Tasks communicate using two MPI-like primitives: a non-blocking Send() and blocking
Receive(). The main advantage of this approach is that a message is only injected into the NoC if
the receiver requested data, reducing network congestion. To implement a non-blocking Send(), a
dedicated memory space in the kernel, named pipe, stores each message written by tasks.
According to [TAN06], a pipe is a communication channel where messages are consumed in the
same order that they are stored. Within this work, the pipe is a memory area of the kernel reserved
25
for message exchanging, where messages are stored in an ordered fashion and consumed
according it. Each pipe slot contains information about the target/source processor, task
identification and the order in which it is produced.
Figure 7 illustrates a communication between two tasks, A and B, mapped at different PEs.
When task A executes a Send() (1), the message is stored in the pipe A, and computation
continues (2). This operation characterizes a non-blocking writing. If the same processor contains
both tasks (sending and requesting messages), the kernel executes a read directly in the pipe. In
this example, task B mapped in other processor, executes a Receive() (3), and the kernel sends a
message request through the NoC (4) to the sender. The scheduler puts task B in wait state (5),
characterizing a blocking read. When task A receives the message request (4), the kernel set the
pipe slot of the message as empty (6) and sends the message through the NoC. When the
message arrives at task B (7), the kernel stores the message in the task B memory space, and
task B returns to a ready state (8).
This communication protocol is susceptible to fail in the presence of faults. For example, if a
fault is detected during the message delivery, task A will fill the communication path with flits, and
will be blocked due the fault. Then task B will wait indefinitely for the remaining of the packet,
ultimately blocking the application.
TaskA
Send()
2 pipe[i].status=USED
pipe[i].status=EMPTY
6
TaskB
1
Receive() 3
TaskStatus=WAITING
4
e st
Requ
g
s
M
MsgDe
5
livery
7
TaskStatus=READY
8
Figure 7 – Communication protocol adopted in the reference MPSoC.
Figure 8 expands the previous Figure to illustrate basic concepts used to measure latency
in this text:
•
Network latency, time spent on a single packet transfer, from the moment it is injected into the
network hardware until it reaches its destination. This time is typically measured in tens to
hundreds of clock cycles.
•
Protocol latency, time to complete a packet transaction. This transaction may consist of
multiple packet transfers, thus, the protocol latency consists of few network latencies, plus the
time spent by the kernel to handle each packet transfer. This measure is typically few
thousands of clock cycles.
•
Application latency, time between two transactions, including the application computation time.
Each type of latency is subject to several sources of variability. For instance, network
latency is subject to network congestion, protocol latency is subject to software interruption time
and CPU load, and the application latency is subject to the amount of computation per packet
transaction.
26
Send()
Receive()
est
MsgRequ
MsgDeliv
ery
computation
time
network
latency
protocol
latency
1
2
application
latency
3
computation
time 4
Send()
Receive()
u e st
MsgReq
MsgDeliv
ery
Figure 8 – Basic performance concepts in distributed systems that will be used in this Thesis.
2.2
Basic Fault Tolerance Concepts
To understand how errors can happen in MPSoCs, we first have to understand how they
can happen in integrated circuits. The three terms: defect, error, and fault are related to system
failure and thus need to be defined.
In [BUS05], a defect is defined as: “A defect in an electronic system is the unintended
difference between the implemented hardware and its intended design”. For example, there is a
set of possible common defects in integrated circuits, such as:
•
Process Defects – missing contact windows, parasitic transistors, oxide breakdown;
•
Material Defects – bulk defects (cracks, crystal imperfections), surface impurities;
•
Age Defects – dielectric breakdown, electromigration;
•
Package Defects – contact degradation, seal leaks.
In today’s technology, we can observe a growth in the defects due to aging and
manufacturing process which are caused by the miniaturization of the transistor size [STA11].
Effects that were not relevant in past technologies are happening more frequently today, such as
electromigration (EM) and negative bias temperature instability (NBTI), and even created new
problems such as positive bias temperature instability (PBTI) [MAR13].
The defect can lead to an error, defined as: “A wrong output signal produced by a defective
system is called an error. An error is an “effect” whose cause is some “defect.” [BUS05]. Note that
a defect may or may not lead to an error. For example, a short to ground at an input of a AND gate
do not show an error if the input is zero.
[BUS05] also defines fault as “A representation of a “defect” at the abstracted function level
is called a fault.” Note that the difference between a defect and a fault is rather subtle. They are the
imperfections in the hardware and functions, respectively. Some examples of the failure modes
resulting from these defects are opens, shorts, leakage, voltage threshold shift, variability in
mobility.
The diminishing reliability of very deep submicron technologies is a fact. Moreover, it is
widely accepted that nanoelectronic-based systems will rely on a significantly lower reliability rate
than what was known so far. The probability of these defects is likely to increase with technology
scaling, as a larger number of transistors are integrated within a single chip and the size of the
chips increases, while device and wires sizes decrease.
Reliability and fault tolerance are very important requirements in current and future SoCs
and MPSoCs. A reliable SoC provides detection and recovery schemes for manufacturing and
27
operational faults to increase not only system reliability and dependability, but also yield and
system lifetime [COT12].
2.3
General assumptions related to the architecture and fault model
The present work adopts a set of assumptions related to the platform to receive the
proposed fault-tolerant mechanisms. These assumptions can be divided in the MPSoC model and
the NoC model. This Thesis adopts common features found in MPSoCs [JAV14][GAR13]:
•
Each PE contains at least one processor with a private memory. This limitation is
bind to the FT NoC approach presented in Chapter 4 where the fault-free path
processing requires a programmable PE running the FT software protocol;
•
Applications are modeled as task graphs and the communication model is message
passing;
•
A online mapping function maps tasks onto PEs, being possible to have more than
one task per PE due to a multi-tasking operating (OS) system – this OS is referred
as kernel along the text. The online mapping function is required to efficiently decide
the new location on the MPSoC of tasks affected by faulty PEs;
•
Fault-free task repository memory, i.e., the memory storing the object codes is
assumed protected against errors;
•
The Manager Processor is assumed to faulty-free.
Concerning the NoC features, we adopt the following common features found in NoCs
[AGA09]:
•
2D-mesh topology;
•
Wormhole packet switching;
•
Adaptive routing algorithm;
•
Duplicated physical channels.
Therefore, platforms with similar architectural features may easily adopt the proposed faulttolerant layered approach.
Testing is at the same time, one of the most costly step in a fault-tolerant method and highly
dependent of reliability constraints and availability of an application. For example, chip testing
requirements for automobile or avionics are typically higher compared to consumer electronics.
Thus, the Thesis is focused on the recovery step, providing a choice of any desirable method for
testing and fault detection according to application needs. For this reason, the actual test of PEs
and the NoC are out of the scope of this work. There are recent proposals that could be used to
test the PEs [GIZ11][PSA10] and the NoC [LUC09][FOC15]. This paper assumes the existence of
a test mechanism to detect faults, responsible to isolate the faulty region (by using the test
wrappers presented in Section 3.2) and firing the fault recovery methods.
The present work assumes only permanent faults, detected at runtime. Transient faults
are out of the scope of the present work, and the literature provides several methods to cope with
them (as [LUC09][COT12][WAN06][PSA10][KER14]. Fochi [FOC15] extended the current work to
cope with transient faults, enabling the network to operate in degraded mode.
28
3 PHYSICAL LAYER
Transport Layer
- FT Communication Protocol
- Fault Free Path Storage
- Path Search Triggering
Application Layer
- PE disabling
- Save task internal state
(checkpoints)
- Restore task (rollback)
5
6
A
PE
Network Layer
- Fault Tolerant Routing
- New Path Search
- Path Reconfiguration
4
B
Physical Layer
- Link Fault Detection
- Fault Isolation (Link
Wrapper)
3
This Chapter details the lowest level of the fault-tolerant method, the physical layer. This
layer is responsible for detecting network faults, signalize faults to the neighbor routers, and isolate
the faulty region.
This layer may contain the test mechanism. In this Thesis, the test is out of the scope, then
we present only an overview of some proposals at the physical layer. The objective is to show that
the architecture is ready to cope with a fault detection mechanism. Furthermore, we introduce the
concept of the Test Wrapper module, which will be used in the upper layers.
3.1
Examples of Physical Layer Approaches
Fochi et al. [FOC15] presents a NoC test approach. Figure 9 presents the approach, which
uses CRC decoders to detect faults. The router can receive CRC decoders in the following
locations:
•
Before the input buffer (CRC 1), with the objective to detect faults in the channel.
•
After the buffer (CRC 2), with the objective of detecting faults in the buffer. The channel can be
healthy, but a fault can change the state of a given bit stored in the buffer.
•
At the output ports (CRC 3), for detecting internal errors of the router. When this kind of error
is detected, it is considered as a critical failure. Moreover, in this case, the router should be
disabled because the integrity of the packets cannot be guaranteed.
Channel 0
Data+CRC
Inputbuffer
CRC
Decoder1
Fault in channel 0
Fault in buffer
Router
CRC
Decoder2
Crossbar
CRC
Decoder3
Fault in the
router
Figure 9– Fault-Tolerant Router architecture proposal.
It is very important to highlight that the CRC scheme does not guarantee 100% of fault
coverage in the router. The CRC does not protect the control logic, and additional test mechanisms
should be considered if higher coverage is required. For instance, [CON09] uses TMR in the state
machines responsible for the buffer control.
29
The work in [FIC09] proposes fault tolerance at the physical level, with a mechanism for port
swapping. Figure 10 on the left shows a router with a fault in the south input port and the east
output port. In Figure 10 on the right, we can observe the configuration after the port swapping.
The result is that three ports may be used, instead 2 before the port swapping. Matos et al.
[MAT13] also presents a fault-tolerant scheme with not only the port swapping scheme, but also
with fault isolation and a mechanism enabling to share buffers’ slots.
Figure 10 –Port swapping mechanism [FIC09].
Hébert et al. [HEB11] proposes a distributed fault-handling method for a NoC-based
MPSoC. The Authors employ a heartbeat mechanism to detect faults in PEs. Each PE has a
watchdog timer capable of detecting missing or corrupted heartbeat messages. Besides fault
detection, the method also comprises faulty PE isolation and recovering via task remapping. Their
main contribution is the fault handler service that checks the health status of a neighbor PE. To
establish the health of an isolated core, the method compares the health register of the neighbor
core with the last diagnosis result. Finally, if the PE is considered healthy, the last function is
executed: un-isolate the healthy PE. This work focused mainly on the fault detection layer. Its main
advantage is that transient and permanent faults are considered, since the core can be reused
after a fault. In the present Thesis, once a fault is detected, the fault is considered as permanent.
Such limitation is due to the scope of our work: implement fault tolerance from the network layer up
to the application layer.
3.2
Physical Layer Proposal
Complex SoCs, such as MPSoCs, need test wrappers such that each core can be tested
individually to reduce both test data volume and test time [COT12][MAR09]. The basic element to
build a wrapper around a given IP is the test wrapper (TW) cell. The main function of the TW cell is
to switch between different operation modes. For example, in Figure 11, the three upper control
signals (test, normal_mode, fault) signal controls the operation modes of the TW cell:
•
Normal Mode - the wrapper logic is transparent to the core, connecting the core terminals to
the functional interconnect, i.e., the input func_in is transferred to the output func_out;
•
Test Mode - used for the actual test of the core. It configures the wrapper to access the core’s
primary I/O and internal scan chains via the test access mechanism.
In the scope of this Thesis, the main idea is to use existing test wrapper cells for the
purpose of fault isolation. In order to support this, the TW cell includes an extra multiplexor, as
illustrated in Figure 11(a) in dark gray, controlled by the fault signal. Activating the fault signal, the
outgoing data of the TW cell (func_out) assumes the binary value ‘V’ connected to the multiplexor.
30
With such approach, it is possible to isolate signals that are interconnecting two modules, avoiding
faults propagations. The idea is that the test detection mechanism controls this fault signal,
activating it if a fault is detected. The Figure 11(b) shows the locations of the TW in the router.
Note in Figure 11(b) that only input ports receive test wrappers. The reason is that the most
complex module of the router is the input port, which comprises the input buffer and its control
circuitry. The TW at the input port avoids the propagation of incoming data to other routers.
normal_mode
fault
test
North Port
V
func_in
Local
Port
TW
TW
T
TWW
East
port
func_out
scan_in
D Q
scan_out
ch 0
West link 0 ch 1
port link 1 ch 0
ch 1
Crossbar
(10 x 10)
TW
TW
TW
TW
Arbitration
and Routing
TW
TW
clock
South port
(a) Test Wrapper (TW)
(b) TWs locations at the router
Figure 11 – In (a) the architecture of the test wrapper cell (TW), with an isolation mechanism
controlled by the fault signal. In (b) the location of each TW at the router
The granularity of the isolation is an important design choice. The designer may:
•
Isolate completely the PE, including the IP and the router;
•
Control the isolation of the IP and the router separately;
•
Control the isolation of the IP and the routers’ links separately;
•
Control the isolation of the IP and the routers’ channels separately.
The main drawback of the first option is to disable the whole PE, even if it is healthy, when
the fault is detected only the IP, but not the NoC. The second choice may lead to an unreachable
IP, even if it is healthy, when the fault is detected at the router.
A trade-off that supports graceful performance degradation is to control the isolation of the
faulty regions at the routers’ ports, links or channels, individually. As the router has ten input
channels, an individual fault signal may be used per channel. By abstracting the test method that
detects the faulty modules, the designer choose the most appropriate test method, trading-off fault
coverage, test time, and silicon area overhead.
In this Thesis, the isolation is tightly coupled with the FT NoC routing, which do not support
a path search at the channel level. For this reason, when a fault is detected in the channel 0 of
north link, for example, the whole north link is disabled (both channels). [FOC15] extended the
current work disabling each link individually, making the channel to operate in “degraded mode”,
where packets are routed to the channel 1, if the channel 0 is faulty (and vice-versa).
Therefore, this Thesis adopts isolation of the links separately. Even if one channel is
healthy, the link is completely disabled. Instead of having ten fault signals for disabling each
channel, there are have five fault signals, one for each link.
Figure 12 illustrates the location of the TW cells between two routers. Note in Figure 12
that, for fault isolation purposes, TW cells are required only in the control signals (tx and credit).
The tx signal signalizes incoming data. The output of the TW cell connected to tx signal, in the
31
presence of a fault in the input link, is ‘0’ (disabled). This action avoids the reception of new data by
the faulty link (Figure 12 – 1), and consequently its propagation to other routers. The credit signal
signalizes that the input buffer has space to receive new flits. The output of the TW cell, connected
to credit signal, in the presence of a fault in the input link, is ‘1’ (Figure 12 – 2). This is a very
important action since it enables the healthy router to drain the buffer connected to the output link
transmitting flits to the faulty router. In this way, all flits transmitted to a faulty link are discarded, by
keeping the credit signal enabled. The upper protocol layers will retransmit the discarded flits using
a different healthy path, as described in the subsequent Chapters.
fault
TW
fault
TW
ROUTER A
2
credit TW
data
TW
tx
eop
ROUTER B
fault
TW
fault
1
TW
Figure 12 – Test Wrappers controlled by a fault signal.dfa
The IP connected to the two local links is responsible for writing in the local link fault
signals. Therefore, by activating the fault mode of both local links corresponds to the isolation of
the IP connected to the router, corresponding to a fault in the processor or another module of the
processing element.
By activating a given fault signal of the East/West/North/South links, it is possible to
degrade the router’s performance gracefully. This means that the router may still be used even if a
given input link is faulty. If the router BIST detects faults in other modules, as the routing and
arbitration logic, it may activates all five fault signals, completely isolating the PE to the rest of the
MPSoC. Such complete isolation is necessary since the router is not able to transmit packets
correctly.
TWs at the input links are sufficient to avoid fault propagation. Thus, it is not necessary to
include TW at the output links. Figure 13 presents the location of the TW cells, as well as the
connection between neighbor routers.
N0 N1
N0 N1
L1
L0
E0
W0
TW
W1
TW
TW
IP
TW
TW
TW
IP
TW
TW
TW
Crossbar
(10 x 10)
TW
TW
E1
L1
L0
E0
W0
TW
W1
TW
Crossbar
(10 x 10)
Arbitration
Arbitration
and Routing
TW
S0 S1
TW
TW
E1
and Routing
TW
TW
TW
TW
S0 S1
Figure 13 – PE architecture, detailing the internal router structure (TW: test wrappers).
32
3.3
Final Remarks
This Chapter presented the physical network level infrastructure that uses test wrappers to
isolate faulty links. One important feature of the proposed method is the fact that the fault isolation
is decoupled of the fault detection. In this way, the designer is free to choose any method
available in the literature that is best suited to the design’s constraints. Moreover, the designer is
free to choose where the fault detection modules can be inserted.
This Chapter presented (according to Figure 14(a)) the option to cope with faults in three
different locations:
•
fault in the output port due to for example for some error in the crossbar of the left router;
•
fault in the link itself (wires), e.g., crosstalk;
•
fault in the input buffer, e.g. SEU (Single Event Upset) in some flip-flop of the input buffer.
The TW cell can isolate the link, avoiding the transmission of corrupted data. Figure 14(b)
represents how the physical fault is represented graphically along this Thesis.
Crossbar
(10 x 10)
TW
Crossbar
(10 x 10)
TW
E1
TW
Arbitration
Arbitration
and Routing
and Routing
TW
output
link
link
TW
TW
TW
input link
(a) possible fault locations
TW
W1
PE
TW
PE
(b) Abstract view of a fault
between two PEs (IP+router)
Figure 14 – Inter PE architecture, detailing the interconnections and fault locations in the router.
This isolation method requires minimal extra hardware and can be adapted to other router
designs with different number of links. Moreover, it enables to reduce gracefully the network
performance as the number of faults increases since it is possible to shut off one router link
individually. The isolation method also ensures that the defective area does not disturb its vicinity,
and any incoming packet is automatically dropped, avoiding fault propagation. Finally, the fault
detection method (BIST circuitry) is decoupled from the rest of the fault recovery approach. The
selected test method has just to write in the fault register when it detects a defective router link.
33
4 FAULT-TOLERANT ROUTING METHOD – NETWORK LAYER
Transport Layer
- FT Communication Protocol
- Fault Free Path Storage
- Path Search Triggering
Application Layer
- PE disabling
- Save task internal state
(checkpoints)
- Restore task (rollback)
5
6
A
PE
Network Layer
- Fault Tolerant Routing
- New Path Search
- Path Reconfiguration
4
B
Physical Layer
- Link Fault Detection
- Fault Isolation (Link
Wrapper)
3
This Chapter presents the state-of-the-art in the field of fault-tolerant NoC routing
algorithms. Then, it presents the second layer of our FT MPSoC, which consists of the proposal of
a set of requirements for FT routing algorithms and a novel FT routing method for NoCs.
4.1
State-of-the-Art in Fault-Tolerant Routing Algorithms for NoCs
Homogeneous NoC-based MPSoCs are a particular example of SoCs with natural
hardware redundancy, with multiple identical processors interconnected by multiple identical
routers. Therefore, if some processors or routers fail, tasks assigned to these processors may be
migrated to other functional processors, and if a router fails, a new path in the NoC may be used.
The Authors in [RAD13] survey the failure mechanisms, fault models, diagnosis techniques, and
fault tolerance methods in NoCs.
Common approaches for fault-tolerant NoC can be divided in two categories:
(i)
Add hardware redundancy to the NoC to deal with faults. Recently, Cota et al. [COT12]
surveyed this category, which includes methods such as ECC and CRC codes, data
retransmission, spare wires, TMR, backup path, and spare routers;
(ii) Add logic to enable the NoC using its natural path redundancy to find fault-free paths.
The first category is typically NoC-dependent, adding more hardware than the second
category. This category can be categorized as static NoCs, with their parameters and structures
fixed at design time [SHA14]. However, a general static NoC-based MPSoC cannot be efficient
due to the dynamic behavior of applications running in MPSoCs, and the high probability of
occurring faults in the chip during its lifetime. Our approach is classified in the second category,
resulting in “dynamic” NoCs, due to its ability to reconfigure itself, avoiding faulty paths.
4.1.1 FDWR [SCH07]
Schonwald et al. [SCH07] propose an algorithm using routing tables at each router. The
Force-Directed Wormhole Routing (FDWR) stores the number of hops to each destination for each
port. This information enables to compute not only the shortest, but also all possible paths to all
destinations. The algorithm updates the routing tables each time the network injects a packet.
Before sending any packet, the router asks to all its neighbors the number of hops to the
destination. Then, it chooses the neighbor with the minimum hop number and updates its table.
Each router executes this process. Thus, the average time to send packets varies between 50 and
34
100 cycles, which is too long compared to non-fault tolerant NoCs (2 to 5 clock cycles). The
Authors present results for 2D mesh and torus topologies. Due to the use of tables, any topology
may be supported. The drawback of the approach is scalability since the table size increases with
the NoC size.
4.1.2 Flich et al. [FLI07]
A similar approach to reduce the size of the tables, by dividing the NoC into virtual regions,
was proposed by Flich et al. [FLI07]. The Authors propose a region-based routing mechanism that
reduces the scalability issues related to the table-based solutions. The table-based routing requires
an input port and the set of nodes associated with a given input – output port.
In a region-based approach, the idea is the combination of as many routers as possible in a
group that shares the same routing path. Then, instead of having one table for each router, they
can reduce the number of elements in the table and consequently reducing the area and
performance overhead.
For example, in Figure 15, some routers have been grouped into regions (r1, r2, r3). The
algorithm defines regions for the reference router (highlighted with a circle). All packets arriving in
the reference router addressed to any destination included in region r1 necessarily need to use the
W output port at the reference router.
From an initial topology and routing algorithm, the mechanism groups, at every router,
destinations into different regions based on the output ports. By doing this, redundant routing
information typically found in routing tables is minimized. Figure 15 shows a region grouping (r4
and r5).
Figure 15 - Example of region definitions [FLI07].
The Authors divide the approach in hardware and software implementations. The first one is
responsible for computing the faulty links of neighbor routers, and the route options using this
information. After each router has computed its connections and possible routes, the software part
can be applied to group different regions.
The Authors divided the algorithm for computing regions into sequential phases. In a first
phase, the algorithm computes the possible set of routing paths for every source-destination. In the
second phase, the algorithm computes the routing regions from the routing options. At every router
the algorithm groups reachable destinations through the same output ports on the router and
coming from the same set of input ports.
35
The Authors present throughput results in different faulty scenarios for different NoC sizes
with a 2D mesh topology. As a conclusion, the region-based approach reduces the overhead
induced by table’s size but do not eliminate it.
4.1.3 FTDR [FEN10]
Feng et al. [FEN10] present a fault-tolerant routing algorithm, named FTDR, based on the
Q-routing approach [BOY93], taking into account the hop count to route packets in the network.
This table-based routing algorithm was modified to divide the network into regions, resulting in the
FTDR-H algorithm. Compared to FTDR router, FTDR-H router reduces up to 27% the NoC area.
When compared to a non-fault tolerant router, for a 4x4 NoC, the FTDR-H area overhead reaches
100%. Moreover, the number of tables increases with the NoC size. The Authors do not discuss
the number of maximum faults that the algorithm supports, the supported topologies and
reachability.
4.1.4 uLBDR [ROD11]
Rodrigo et al. [ROD11] present the uLBDR. The first version of the algorithm, LBDR
[ROD09a], stores the connectivity status of the neighbors’ routers (Ceast,Cwest, Cnorth and Csouth) and
the turns (Reast-west, Reast-north, etc.) to reach its destination. It uses an adaptive routing for reaching
the destination, using the connectivity status to use another path.
The problem with this approach is that it only ensures minimal path, leading to scenarios
with unreachable routers. Then, if the neighbors’ information is not sufficient to reach the
destination, a deroute process (algorithm named LBDRdr) is executed, trying to route the packet in
a non-minimal path. However, even with the deroute process full reachability is not ensured.
The Authors propose a third mechanism that ensures complete reachability of routers: the
replication of the packets in two output ports: uLBDR. The drawback is that packet replication
potentially increases the network congestion and a virtual cut through (VCT) approach is applied to
avoid deadlocks. VCT requires larger buffers to store the entire packet and has longer network
latency compared to wormhole.
The results show that compared to LBDR, the uLBDR increases 44.7% the silicon area.
They also present results regarding power, frequency and flit latency. The presented results are for
2D irregular meshes, not exploring other topologies.
4.1.5 iFDOR [SEM11]
Sem-Jacobsen et al. [SEM11] propose two modifications in the original FDOR (Flexible
Dimension Ordered Routing) algorithm [SKE09]: (i) create a boundary around a faulty router; (ii)
reroute packets around the faulty router. Figure 16 shows a scenario where a faulty router is
present (white router). The boundary around the faulty router is made reconfiguring its neighbor
routers’ as boundary routers. This way, any packet is routed around the faulty router.
Two virtual channels are used to ensure deadlock freedom by prohibiting one turn. The
results have shown an increase of 279% in terms of area for a mesh NoC. The Authors do not
discuss the reachability feature and the number of maximum faults that the algorithm may support.
Moreover, a single fault disables the entire router, quickly reducing the network performance as the
number of faults increases.
36
Figure 16 - The path followed by a packet around the fault (blue and red lines). The white node is the
exclusion region, the gray nodes are the reconfigured as boundary routers.
4.1.6 Vicis [DEO12]
DeOrio et al. [DEO12] and Fick et al [FIC09] also propose fault-tolerant routing algorithms
using tables. All routers contain a table to all router’s destinations. The algorithm is divided in two
steps. First, each router selects the direction to route a given destination based on its neighbors,
updating the entry in its own table. Second, the router applies a set of rules to avoid deadlock.
These rules disable turns without requiring virtual channels or duplicated physical links. On the
other hand, this approach does not ensure full reachability. Presented results for mesh and torus
topologies guarantee that at least 99,99% of paths are deadlock-free with up to 10% of faulty links.
4.1.7 Alhussien et al. [ALH13]
The Authors propose a fault-tolerant routing algorithm for NoCs. The algorithm is built
based in a topology named DMesh [CHI11]. The topology is based in a mesh architecture, but the
originality is a connection in the diagonal between routers as shown in Figure 17.
Figure 17 – DMesh topology.
The Authors developed a routing algorithm for supporting up to three faulty channels. The
algorithm is based in the information of neighbor faulty links to avoid the faulty channels. They
presented results for average latency for one, two and three faults for different traffic rates and
patterns. Because the modifications are only made in the routing logic, the increase in router’s area
is only 7%. It is important to highlight that with more than three faults in the NoC the routing
algorithm presents deadlock occurrences.
37
4.1.8 MiCoF [EBR13]
The method named MiCoF (Minimal-path Connection-retaining Fault-tolerant approach) is a
combination of routing algorithm with a hardware support in the presence of faults. In the hardware
scheme, the router connections are modified with the objective to maintain connectivity in the
presence of faults. As shown in Figure 18 the east input channel is directly connected to the west
output channel while the west input channel is connected to the east output channel. Similarly, the
packets coming from the north or south input channels are directly connected to the south or north
output channels, respectively. A similar approach was also taken by [KOI08].
The second contribution is a fault-tolerant routing algorithm to this architecture. This
proposal takes into account the neighbor router fault information to route the packets. When a
router is faulty, the links are simply connected to each other along the horizontal and orthogonal
directions. Using hardware bypass (EàW, SàN, etc.) the Authors propose modifications in the
routing algorithm. The route taken is modified according the fault location and destination. Figure
19 presents some scenarios with up to two faulty routers.
Figure 18 – Connection modification to maintain connectivity in presence of faults.
The algorithm showed a reachability of 99.998% in all combination of two faults in a 8x8
mesh topology, but maintaining the minimal path. Increasing the number of faults reduce the
reachability percentage of the algorithm. The Authors showed that for six faults, MiCoF can
guarantee 99.5% of reachability. The MiCoF area for a UMC 90nm technology is 7.295 mm².
Figure 19 – MiCoF Routing in scenarios with two faulty routers [EBR13].
38
4.1.9 BLINC [DOO14]
The Authors propose modifications in a baseline router. Initially, routing tables are
generated offline with all the routing is based on them. If a fault affects the network topology,
BLINC utilizes a combination of online route computation procedures for immediate response,
paired with an optimal offline solution for long term routing.
When the BLINC reconfiguration is triggered, two actions are started at parallel:
(i)
a complete routing reconfiguration in software to generate new minimal routing tables for
each router;
(ii)
enabling pre-computed routing metadata, so to quickly resolve the affected routes. The
rerouting response from BLINC is deadlock free, but not necessarily minimal.
To quickly find emergency routes upon a failure, while the complete routing reconfiguration
is executed, BLINC employs compact and easy-to-manipulate routing metadata. The routing
metadata consists in a region-based table (in the paper the term is segments). While the
reconfiguration is executed, the routers within this region use this region-based table to route
packets.
The evaluation shows more than 80% reduction in the number of routers affected by
reconfiguration, and 98% reduction in reconfiguration latency, compared to state-of-the-art
solutions. There is no discussion on the area overhead. The Authors only state that an 8x8 mesh
requires at least 264 bits per router to store the metadata.
4.2
Discussion on the state-of-the-art
Evaluating the proposed works, we enumerate below features that should be present in a
fault-tolerant routing algorithm, besides being deadlock and livelock free:
(i)
Generic: being possible to adapt it to different NoCs, including different NoC topologies;
(ii)
Complete reachability: i.e., if a path between a given source-target pair exists, the routing
algorithm must find it;
(iii)
Granularity of the fault location and the moment it occurs: i.e., faults may reach routers
and/or links at any moment. Some routing algorithms restrict the number of simultaneous
faults [DOO14] or if the fault occurs in links or routers [SEM11];
(iv)
Scalable: the cost must be independent of the NoC size. Table-based approaches
(discussed in the state-of-the-art) have their area cost linked to the NoC size;
(v)
Local path computation: i.e., algorithms with a PE responsible to store and compute all
paths must be avoided due to the large size of actual SoCs (a feature also related to
scalability). Each router or PE must be able to compute the new path if a fault is detected,
without interacting with a central manager, which can also be subject to faults;
(vi)
Acceptable cost: i.e., the implementation of the proposal routing algorithm in a reference
NoC should not severely impact the silicon area, power, and its performance.
We divide the FT NoC routing algorithms in two categories: (i) table-based approaches and
(ii) distributed. In the first category, each router contains a table with two entries: target router and
port. When receiving a packet, the routing module searches the port to forward it. This approach
fits well with fault tolerance purposes because the table can be easily modified, changing the path
to be taken to a given destination. The main advantage of table-based routing is that we can use
any topology and any routing algorithm. The problem is that this approach is not scalable since
routing tables are implemented with memories, also does not scale in terms of latency, power
39
consumption, and area. Thus, this approach is unfeasible for large NoCs [ROD11]. The number of
destinations in this table is a function of the NoC size: a 3x3 NoC each router has nine destinations
while in a 10x10 NoC there are 100 possible destinations. Some proposals for this approach can
be found in the state-of-the-art [SCH07][DEO12][FIC09]. To solve this issue, some works propose
to divide the NoC in virtual regions [FEN10][FLI07]. Then, instead of having a table with all possible
destinations, some routers are grouped in “regions”, reducing the number of destinations in the
table. For example, Figure 20 shows the combination of routers in regions proposed by Flich et al.
[FLI07]. Unfortunately, the main drawback of such mechanism is that, even with 16 regions, it still
does not achieve full reachability and induces a long critical path [ROD09b].
Figure 20 – Region computation process proposed by [FLI07].
Table 2 compares the reviewed approaches, positioning our method to the evaluated
features. The closest approach is the uLBDR method, however the duplication of packets affects in
the network traffic, increasing with the NoC size and needing a packet dropping method for the
duplicated one. The fault-tolerant routing algorithms are a function of the NoC topology, as mesh
and torus. Flich et al. [FLI12] evaluate topology-agnostic fault-tolerant routing algorithms, not
specifically designed for NoCs, but the Authors claim that they may be used for NoCs. They
present results regarding average routing distance, computing complexity, average packet latency,
and throughput for mesh and torus topologies, with or without failures. However, results concerning
how much time the algorithm takes to find the path are not presented. They do not present results
for full reachability and silicon area too.
Table 2 – Comparison of different routing approaches.
Proposal References
uLBDR
[ROD11]
iFDOR
[SEM11]
FDWR
[SCH07]
FTDR
[FEN10]
Flich et
al.
[FLI07]
Vicis
[FIC09]
[DEO12]
Routing
Area
Topology
Full
Fault Model
Scalable
Type
overhead
agnostic? Reachability
46%
(compared to
irregular
distributed links/routers
yes
yes
the baseline
mesh
router)
~279%
irregular
distributed
router
compared to
yes
N/A
mesh
FDOR
table links/routers
N/A
no
yes
N/A
region
~100% in
based links/routers
no
N/A
N/A
12x12 NoC
tables
region
240 gates in
yes (practically
based links/routers
no
N/A
8x8 NoC
unfeasible)
tables
300 gates per
irregular
table links/routers router (4x4),
no
no
mesh /torus
330 (12x12)
Alhussien
[ALH13] distributed links/routers
et al.
MiCoF
[EBR13]
BLINC
[DOO14]
Proposed [WAC12b]
Work [WAC13a]
7%
7.295 mm² for
distributed routers a UMC 90nm
technology
table
links
N/A
42% in LUTS/
58% in FFs
path search links/routers (wrt to the
baseline
router)
max #
Offline
faults Computation
Packet
Dropping
VC/Physical
Channel
Minimal
Routing
any
yes
yes
no
yes
N/A
no
no
2 VCs
N/A
N/A
no
no
no
N/A
N/A
no
no
no
N/A
N/A
yes
no
no
no
N/A
yes
no
no
N/A
yes
no
no
3 links
no
no
2 physical
vertical links
yes
no
no
no
3 links
no
no
2 physical
vertical links
N/A
no
N/A
yes
any
yes
no
N/A
yes
yes
yes
yes
any
no
no
2 physical
yes
40
4.3
Proposed Approach: MAZENOC Routing Method [WAC12b][WAC13a]
This section describes the first original contribution of this Thesis: a novel fault-tolerant
NoC routing method. The method is inspired in VLSI routing algorithms [LEE61], and has the
objective to find a path between source-target pairs in faulty networks. The method is generic,
allowing abstracting the network topology. The method requires the following common features
found in NoCs: wormhole packet switching; adaptive routing algorithm; duplicated physical
channels. Features related to fault tolerance include:
(i) Deadlock avoidance mechanism;
(ii) Simultaneous support to both distributed and source routing;
(iii) Isolation wrapper cells at the NoC’s input ports (Chapter 3);
(iv) Presence of a secondary path discover network (Section 4.3.1).
The use of duplicated physical channels ensures deadlock avoidance. The number of virtual
or replicated channels required to avoid deadlocks is a function of the network topology. For
example, two virtual or replicated channels are sufficient to avoid deadlocks in a 2D-mesh topology
[LIN91]. Other topologies such as 2D-Torus and hypercube need additional channels [LIN91].
In the presence of faults, the path to circumvent faulty regions may require a turn that could
lead to a deadlock. Therefore, a fully adaptive routing algorithm is required. Using the replicated
channels, a 2D-mesh NoC may be divided into two disjoint networks, each one implementing a
partial adaptive routing algorithm, for example, west-first and east-first [GE00], resulting in a fully
adaptive routing. In fault-free scenarios, the distributed XY routing is adopted, and when faults are
detected, source routing is applied using west-first and east-first partial adaptive routing algorithms
to reach the destination.
This layer (Network layer) assumes that a higher-level layer, e.g. transport layer, triggers
the proposed path discover procedure when a fault is detected. This can be done with the help of
a fault-tolerant communication protocol or with hardware modules used to detect lost packets, as
discussed in the next Chapter.
At the system startup, the network is assumed faulty-free, and packets are sent from the
source PE to the target PE using the XY routing algorithm. The method, executed at runtime,
comprises three steps, as illustrated in Figure 21: (i) seek new path; (ii) backtrack the new path; (iii)
clear the seek structures and compute the new path. At the end of the described process, the
source PE receives a valid healthy path to the target PE. This path is stored in the PE memory and
the following packet transfers (from this source to the target) use this path. Thus, the proposed
method is executed only once for each missed packet, without requiring the whole process for
each packet transfer.
In the first step of the method, seek step, if the source PE identifies that it is not able to
transfer packets to a target PE, this PE generates a seek request, asking for a fault-free path to the
target PE. The router stores this request into an internal memory containing the fields S/T/P/#,
meaning source router, target router, incoming port, and hop counter. Each router checks if it is the
target router. If not, it increments the hop counter and broadcasts the seek request to its neighbors
(except for the port that originated the request), via a separate network optimized for broadcast.
The next routers store the seek request in their own internal memories, checking if they are the
target router and repeating the process until the target router is reached or a timeout mechanism at
the source PE is fired. In this case, the source PE broadcasts a clear command to free the seek
memory and it tries another seek a random time later. Therefore, if a seek process is blocked due
41
to an overflow in the number of memory entries, it can be executed later. The target router is
declared unreachable after a given number of unsuccessful seek requests.
n
Target router
unreachble
Transmit the
packet
SEEK
BACKTRACK
y
inject backtrack
header
source PE
generates seek
request
mask payload
store entry in the
memory
target router ?
start timer
n
timeout ?
y
n
route packet to
incoming port
y
n
source router ?
y
increment hop#
identify invalid
turns
broadcast to
output ports
broadcast
clear
store new path
in PE
CLEAR & COMPUTE
Figure 21 – Flowchart of the routing method.
The memory size is constant for the NoC design, regardless the NoC size, ensuring
scalability. The number of the memory entries is only related to the maximum number of
simultaneous seek requests, and its size is not proportional to the NoC size, as in the table-based
routing methods. If the number of simultaneous seeks is larger than the memory size, the last seek
waits the end of a seek request to restart the process. If the designer knows that it is unlikely to
detect multiple faults at the exact same time, then it is possible to reduce the memory size to one.
In other words, the designer can tradeoff between the time to determine a new fault-free path and
the silicon area overhead.
Figure 22 illustrates a scenario with four faulty routers, one source, and one target router.
Initially, the source PE requests a seek to its router, creating an entry in its local memory with value
S/T/L/0, meaning that the source request came from the local port (L) and it is hop 0 (Figure 22(a)).
Since this is not the target router, it broadcasts the seek request to the north and east ports.
However, the east port points to a faulty router, which means that its input ports drain any incoming
packet in order to avoid fault propagation, as explained in the previous Chapter. This way, only the
seek request sent to the north port is propagated. The second router (Figure 22(b)) stores the entry
S/T/S/1 in its local memory, meaning it received the request from the south port and it is the hop
number 1. The broadcast continues until it reaches the target router (Figure 22(d)) via the east port
with 8 hops.
42
Faulty
router
Faulty
router
Faulty
router
S/T/L/0
Faulty
router
Faulty
router
SOURCE
(a)
S/T/L/0
Faulty
router
...
Faulty
router
SOURCE
S/T/L/0
Faulty
router
Faulty
router
(b)
S/T/S/2
Faulty
router
Faulty
router
SOURCE
TARGET
Faulty
router
S/T/S/1
S/T/S/1
S/T/S/3
Faulty
router
S/T/S/5
S/T/S/6
S/T/S/2
S/T/W/3
S/T/W/4
S/T/W/5
S/T/S/1
Faulty
router
Faulty
router
S/T/N/6
S/T/L/0
Faulty
router
S/T/E/8
S/T/N/7
SOURCE
TARGET
TARGET
(c)
TARGET
(d)
Figure 22 –Seek steps of the method in a 4x4 mesh, with four faulty routers.
The second step, backtrack step is executed only once, and starts when the target router
is reached by the first received seek. All subsequent seeks are discarded. This feature ensures
that, if there is only a single seek request at the moment in the network, the shortest path to the
target is found. The target router injects a backtrack packet into the NoC, where its payload
correspond to the ports selected by each router to reach the target router. The backtrack packet
uses source routing, following the reverse path followed by the seek step. The backtrack ends
when the source router is reached and the backtrack packet is sent to the source embedded
processor to check the validity of the path.
Figure 23 shows the backtrack process, where the target router creates the backtrack
packet. Its payload is initially filled with don’t care data. At the first router (the target router), it
inserts its seek incoming port (E) into the payload. Then, the packet follows the direction stored
into the internal memory (E), reaching the next router. The next router also inserts its incoming port
(W) into the payload. Then, the packet follows the direction stored into the internal memory (S),
reaching the next router via its south port. In this way, the backtrack packet follows the reverse
seek path, until it reaches the source router where the payload is now fully filled with the path to
the target router. The path in this example is [N N E E E S S W E].
S/T/S/3
Faulty router
S/T/S/5
S/T/S/6
S/T/S/2
S/T/W/3
S/T/W/4
S/T/W/5
[--EEESSWE]
S/T/S/1
[---EESSWE]
Faulty router
[----ESSWE]
Faulty router
[NNEEESSWE]
SOURCE
S/T/N/6
[-----SWE]
[-NEEESSWE]
S/T/L/0
[----SSWE]
Faulty router
S/T/E/8
S/T/N/7
[-------E]
[------WE]
TARGET
Figure 23– Backtracking process of the proposed method.
43
The backtrack packet may not be received in two cases: the target router is unreachable, or
the number of maximal simultaneous seeks is exceeded. In the second case, the source router can
initiate a new seek after a random time. If all the seek requests failed to receive backtrack, then the
target router is declared unreachable.
The third step is named compute path and clear seek. The path followed by the backtrack
packet might create a deadlock. For this reason, the source embedded processor checks the path,
for instance [N N E EE S S W E], for invalid turns. When there is an invalid turn, it forces the
packet to change the physical channel (or virtual channel), breaking the dependency cycle. For
instance, the path presented above has a turn from south to west. This path should be changed
from channel 0 to 1 because it is a prohibited turn in the west-first routing algorithm, but allowed by
east-first one. Thus, the corrected path to the example target router must be [N0 N0 W0 W0 W0 S0
S0 W1 E1]. Once this path is computed, it is stored in the PE main memory for future packet
transmissions.
In parallel to the path computation, the source PE requests to clear the entries in the seek
memory labeled with the source router address. This clear command is broadcasted to the entire
seek network, similarly to the seek process.
Note that the path computation depends on the network topology and its routing algorithms.
For this reason, this part is implemented in software such that it can be easily modified. The
hardware structure, presented in next section, is not dependent on the network topology and the
routing algorithms.
4.3.1 Hardware Structure
Broadcasting messages in the NoC would create a congested scenario, disturbing the NoC
performance. For this reason, a seek routing module was created, separating the NoC from the
“seek network”. Figure 24 illustrates how the seek routing module is connected to its neighbor
routers (in this example they are N, S, W, E) creating two separated networks. The conventional
NoC links are represented by dashed lines (number 2 in Figure 24) and the seek links are the
continuous lines (number 1), consisting of 26 wires in this example. It is worth to mention that the
seek links and the seek module are used during the seek/clear steps to broadcast the seek/clear
commands without disturbing the NoC traffic (Figure 21), while the backtrack step uses the NoC.
PE
router
seek
routing
1
Rs
8
Rt
8
hop#
6
seek
2
N
1
clear
ack_seek
PE
ack_clear
router
W
E
seek
routing
S
L
Figure 24 – Inter-router connection of the Seek router.
44
Note that the local port in Figure 24 has only the incoming seek link because, according to
Figure 21, when a source PE detects an unreachable PE, it starts the seek process. The outgoing
seek link is not required because the target PE does not receive data from the seek module.
Figure 25 details a generic router with 2 physical channels in the local input port. It also
presents the logic required by the proposed method (gray area) and its intra-router signals. The
inter-router links, presented in Figure 24, are not presented here for sake of simplicity. The routing
and arbitration module is slightly modified to interface with the seek routing module. At the
beginning of the backtrack step, the routing and arbitration module identifies a backtrack packet (1)
and asks the seek module to search its internal memory (2) for the corresponding output port for
the current seek request. The seek module answers with the appropriate output port (3). For the
backtrack step, the payload receives the path information stored in the seek memory, through the
added mux (4 and 5).
routing and
arbitation
local port
1
2
4
3
seek
routing
5
input
ports
crossbar
output
ports
Figure 25 – Intra-router connection of the Seek router.
4.4
Results
In this section, the proposed routing method is evaluated in different NoC topologies. We
also evaluate the Time to Find Alternative Paths (TFAP) and the area overhead induced by the
modifications.
4.4.1 Experimental Setup
According to Salminen et al. [SAL08], from 66 NoC designs reviewed, the most common
NoC topologies are mesh and torus (56%), bus and crossbar (14%), custom (12%), fat-tree (11%),
and ring-based (7%). Since bus, crossbar, and fat-tree typically do not provide alternative paths in
the presence of faults, these topologies were not evaluated.
The Spidergon STNoC topology is a regular [COP08], point-to-point topology similar to a
simple bidirectional ring, except that each node has, in addition to links to its clockwise and
counter-clockwise neighbor nodes, a direct bidirectional link to its diagonally opposite neighbor
(Figure 26).
The Spidergon STNoC uses the Across-First routing algorithm. This algorithm moves
packets along the ring, in the proper direction, to reach nodes which are closer to the source node,
and use the across link only once at the beginning for destinations that are far away [COP08]. For
example, a given packet from routers 0 to 5, would take the path 0→8→7→6→5. The Across-First
routing requires two virtual channels to be deadlock free. The Hierarchical-Spidergon (Figure 27)
topology uses two Spidergon networks layers, where each router communicates with its network
and, additionally, with the lower/upper layer.
45
0
1
2
3
4
15
1
2
3
8
9
10
11
7
6
5
4
15
14
13
12
5
14
6
13
12
0
7
11
10
9
8
Figure 26 – Spidergon STNoC, topology and physical view.
0
1
2
3
4
0
1
15
2
3
4
5
15
14
5
6
14
13
6
7
13
12
11
10
7
9
8
12
11
10
9
8
Figure 27 – Hierarchical-Spidergon topology. Red lines show the connection between the lower and
upper layers.
This section evaluates the proposed routing method with four topologies: 2D-mesh, 2DTorus, Spidergon and Hierarchical-Spidergon. All NoCs are configured with 100 routers, with four
routers pairs starting the seek process. Figure 28 shows a configuration for the 2D-mesh and 2Dtorus topologies.
90
80
70
60
50
40
30
20
10
0
91
81
71
61
51
41
31
21
11
1
92
82
72
62
52
42
32
22
12
2
93
83
73
63
53
43
33
23
13
3
94
84
74
64
54
44
34
24
14
4
95
85
75
65
55
45
35
25
15
5
96
86
76
66
56
46
36
26
16
6
97
87
77
67
57
47
37
27
17
7
98
88
78
68
58
48
38
28
18
8
99
89
79
69
59
49
39
29
19
9
Figure 28 – Configuration used in mesh and torus networks. Gray routers are faulty routers and the
dark gray routers are communicating router pairs. Four communicating pairs are illustrated: 81à 9,
45à 47, 41à 49, 9à 55.
For instance, let us assume the communication from routers 81 to 9 using the torus
topology. As illustrated in Figure 29, the proposed method is able to find shortest paths using the
wraparound links 29 to 20 and 0 to 9, resulting in the path SEEEEEENEESSSSSSESSW (steps 1
and 2).
46
90
80
70
60
50
40
30
à 20
91
81
71
61
51
41
31
21
92
82
72
62
52
42
32
22
93
83
73
63
53
43
33
23
94
84
74
64
54
44
34
24
95
85
75
65
55
45
35
25
96
86
76
66
56
46
36
26
97
87
77
67
57
47
37
27
98
88
78
68
58
48
38
28
99 89
79
69
59
49
39
29 à
10 11 12 13 14 15 16 17 18 19 ß 0 1 2 3 4 5 6 7 8 9 ß
Figure 29 – Minimal path found by the proposed method from router 81 to 09 (dark gray), in a torus
topology (gray routers are faulty node). The arrows indicate wraparound links.
The method was also able to find minimal paths in the Spidergon NoCs, assuming the same
four communicating pairs used for mesh and torus. The Figure 30 illustrates the fault scenario for
the Spidergon NoCs. Imagine that we have diagonal connections, e. g. router 0 is connected to 50,
1 to 51, etc. In the detail we have the four source – target pairs of evaluated flows.
0 1 2 3 4 5 6 7 8 9 10111213141516171819202122232425 99
26 98 27 97 28 96 29 95 94 93 92 91 90 89 88 87 86 85 84 83 82 81 80 79 78 77 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 76 49 7574737271706968676665646362616059585756555453525150
9→55
41→49
45→47
81→09
Figure 30 – Spidergon Topology (gray routers are faulty nodes). Diagonal connections were omitted.
In the detail we have the four pairs of evaluated flows
47
4.4.2 Time to Find Alternative Paths (TFAP)
Table 3 shows the time to find paths in the presence of faults for each topology, in terms of
clock cycles, for each step of the method. The values presented for step 3 include only the clean
process, not the path computation. To account for the path computation time, we executed the
path computation algorithm assuming a mesh topology and a MIPS processor running at the same
frequency of the NoC. The result is that the processor takes, on average, 200 clock cycles per hop
to evaluate the turns of the path. The complexity of this algorithm is O(n), where n is the number of
hops in the path.
Next, we evaluate the average results from data presented in Table 3. The average number
of clock cycles per hop (AHcc), assuming all topologies, is 16 and 15.5 for the seek step and the
backtrack step, respectively. It is not required to account for the time of step 3 (clean) since it
happens in parallel with the path computation. Therefore, the AHcc for step 1 and 2 is 31.3
(standard deviation of 2.6). Using these average values, we can estimate that, for instance, a path
of 10 hops would take 2,313 clock cycles, being 313 clock cycles for steps 1 and 2, and 2,000
clock cycles to compute the path. To put this result into perspective, 2,313 clock cycles is a value
inferior to a typical time slice in a multi-task operating system. Therefore, the impact to compute a
new path represents a very small impact in the overall performance of an MPSoC system.
Moreover, recall that, due to the locality of tasks in a MPSoC, two communicating tasks are
expected to be close to each other. Thus, the number of hops is typically smaller than 10 hops.
Table 3 – Time to execute the proposed routing method, in clock cycles, for different topologies.
Topology
Rs→Rt
09→55
41→49
Mesh
45→47
81→09
09→55
41→49
Torus
45→47
81→09
09→55
41→49
Spidergon
45→47
81→09
09→55
41→49
Hierarchical-Spidergon
45→47
81→09
Numberofhops
43
16
16
36
27
16
16
20
5
10
4
29
5
10
4
5
Step1 Step2
Step3
seek
backtrack clean
688
256
256
576
432
256
256
320
80
160
64
464
80
160
64
80
548
221
221
463
354
221
221
269
89
149
77
378
89
154
81
93
635
245
620
530
395
245
440
350
500
155
65
440
246
245
230
245
Once the path is determined by the proposed method, packets are transmitted as standard
wormhole packets, with the same router latency compared to the base router [CAR10], i.e. the
proposed method does not affect the network latency.
The reviewed works do not present similar analysis for path computation, so we can only
perform a rough comparison. For instance, the Vicis NoC [DEO12] takes around 1,000 cycles to
48
reconfigure the network, without accounting the execution of the routing algorithm. In [SCH07] the
average time between the sending of two consecutive packets is between 50 and 100 cycles,
which increase latency. In our method this limitation does not exist, since the path search is
executed once.
4.4.3 Multiple Seeks Evaluation
To evaluate the behavior of the approach with multiple seeks, we analyze a scenario with
four routers simultaneously asking a new path. As an experimental scenario, we consider a 10x10
2-D Mesh, with a set of faulty routers, as presented in Figure 31. This scenario contains one
bottleneck for the seek request at router 42.
99 98 97 96 95 94 93 92 91 90
80 81 82 83 84 85 86 87 88 89
79 78 77 76 75 74 73 72 71 70
60 61 62 63 64 65 66 67 68 69
RS→RT
99→50
78→48
0→49
21→51
59 58 57 56 55 54 53 52 51 50
40 41 42 43 44 45 46 47 48 49
39 38 37 36 35 34 33 32 31 30
20 21 22 23 24 25 26 27 28 29
19 18 17 16 15 14 13 12 11 10
0
1
2
3
4
5
6
7
8
faulty router
target router
source router
9
Figure 31 – Evaluation scenario of a NoC 10x10.
This algorithm returned the expected paths for each pair RS→RT, being all minimal. The
bottleneck happens when the four requesting seeks arrive in the router 42. The seek module
handles each request sequentially. However, each seek will be stored in a different entry in the
seek memory, with the same output (east port). What happens is that each seek wave will find the
minimum path, but the last processed seek will have a seek time (step 1) larger than the others.
In a scenario with five routers simultaneously asking a new path, the last seek request is not
treated since there is no free entry in the memory. In this case, an upper layer should request
another seek at a random time later, when there is space in the seek memory.
4.4.4 Area Overhead
Table 4 presents the area overhead in terms of look up tables (LUTs) and flip-flops (FFs)
compared to the baseline router (with a seek memory with 4 entries). Regarding the LUTs
occupation, the overhead is 42%, and for FFs 58%.
The router was also synthesized using Cadence Encounter RTL Compiler, targeting a 65
nm technology. The router with the seek module required 6,204 cells, occupying 43,049 µm². A
silicon area comparison with previous works is not straightforward because the target router
architecture might be considerably different and the technology library can be different. The best
match is uLBDR [ROD11] router, which is similar to the evaluated router and uses the same
technology. The uLBDR [ROD11] router consumed approximately 62,050 µm² and they do not
present the features of the proposed method. Also, the presented baseline router is very small in
terms of area. Even with the overhead induced by the seek module, the MAZENOC router
(baseline router + seek module) is smaller than uLBDR.
49
Table 4 – Area overhead for Xilinx xc5vlx330tff1738-2 device.
submodule
AreaOccupation LUTs
FFs
module
SwitchControl 351
TOTAL
1610
97
433
baselinerouter
Seekmodule 370
SwitchControl 339
TOTAL
2293
184
157
682
baselineroutermodified+seekmodule
Overhead
58%
42%
Considering that the router represents less than 10% of the PE area, an increase of 50% in
the router area represents less than 5% in the total area overhead of one PE, which is an
acceptable cost considering full reachability even under severe faults scenarios, scalability in terms
of NoC size, no impact on the communication performance after the path calculation, and topology
independency.
4.5
Final Remarks
This Chapter presented an original method for fault-tolerant NoC routing, compliant with the
following features:
1. Topology-agnostic. the approach is divided into a hardware and software parts. The
hardware part, i.e. the seek module, is topology agnostic. The deadlock avoidance
algorithm is topology dependent, but it is implemented in software such that it can
be easily modified. As far as we know, there is no other approach with this feature.
2. Complete reachability: most approaches support a limited number of faults and do
not guarantee complete network reachability. The experiments presented in Section
4.4 demonstrate that the method is able to find a route even with dozen of faulty
routers.
3. Granularity: the fault location can be any port link of any router.
4. Scalability: even though the proposed method uses a small memory, that has a
completely different purpose compared to previous approaches (table-based), such
that its size is not related to the network size.
5. Fast path finding: the method is able to find a new path in few hundreds of clock
cycles per hop. This time is very fast compared to software execution time, context
switching, interrupt handling, etc.
6. No impact on the communication performance: the path finding process uses the
seek network for flooding, instead of the NoC. Once the new path is found, the
packet latency behaves as a faulty free path assuming the same number of hops.
Moreover, the total path search and computation time is small compared to the literature,
with an acceptable area cost. Due to these characteristics the proposed routing method may easily
integrated into another NoC design that require fault tolerance.
Despite the advantages of the proposed method, the transmission of the backtrack packet
uses the NoC to reach the source PE. Therefore, the path taken by the backtrack packet (from the
target to the source PE) may execute invalid turns leading to a deadlock. This issue may be solved
50
adopting store and forward switching technique for these packets. This is possible because even
for large backtrack packets the number of flits is small, being possible to store the entire packet in
the routers’ buffers.
51
5 FAULT-TOLERANT COMMUNICATION PROTOCOL – TRANSPORT
LAYER
Transport Layer
- FT Communication Protocol
- Fault Free Path Storage
- Path Search Triggering
Application Layer
- PE disabling
- Save task internal state
(checkpoints)
- Restore task (rollback)
5
6
A
PE
Network Layer
- Fault Tolerant Routing
- New Path Search
- Path Reconfiguration
4
B
Physical Layer
- Link Fault Detection
- Fault Isolation (Link
Wrapper)
3
This Chapter presents the third layer of the proposed FT communication protocol. This layer
is implemented in the kernel, hidden from the programmer. It triggers the path finding method
presented in the previous Chapter when a message is not delivered. Section 5.1 presents and
discusses the state-of-the-art in FT protocols for message passing. Section 5.2 and the following
ones present original contributions of this Thesis.
5.1
The State-of-the-Art in Protocol and System Reconfiguration Approaches
Fault detection and location (i.e. fault diagnosis) are the first challenges in the quest for
dependable NoC-based MPSoC. This step is not trivial since NoCs may have hundreds of routers.
As surveyed in [COT12], the universe of online test approaches for NoC-based designed includes
the use of error control coding schemes, BIST for NoC routers, and functional testing. These
testing approaches have been extensively studied in the last decade and can be considered
mature. Testing other parts of an MPSoC, such as processors and embedded memories can be
considered even more mature.
The second challenge towards a dependable MPSoC would be the system reconfiguration.
The MPSoC receives the fault location and changes the system configuration, masking regions
with permanent faults. The predominant fault reconfiguration approaches are the use of spare logic
(spare links, routers, or an entire PE), network topology reconfiguration, and use of adaptive
routing algorithm [COT12].
As we move towards a macro architectural view of dependability in MPSoCs, two main
communication paradigms must be taken into account: shared memory and message passing.
Message passing is usually preferred in NoC-based MPSoCs due to the scalability offered by this
communication protocol. Therefore, this Chapter adopts message passing as the inter-task
communication method.
There are several attempts to implement message-passing libraries in MPSoCs, but most of
them are focused on defining an efficient programming model for NoC-Based MPSoC in a layered
communication protocol stack. For instance, Fu et al. [FU10] describe an MPI-based library where
the task mapping is abstracted from the application. It also implements features such as broadcastgather and message ordering. Mahr et al. [MAH08] argue that most message passing libraries for
MPSoCs are specific for a single network. They developed a network-independent library that,
52
according to the target network topology, this library can be simplified or customized, removing the
unnecessary parts and thus reducing the memory footprint.
Fault-tolerant message passing libraries were proposed in the context of fault-tolerant
distributed systems, usually applied to a cluster of computers. An application may have very large
runtime, and faults during this process may corrupt data or stop the application. Aulwes and Daniel
[AUL04] include reliability features on MPI such as checksum, message retransmission, and
automatic message re-routing. Batchu et al. [BAT04] test unresponsive processes by implementing
self-checking threads that use heartbeat messages generated to monitor the MPI/FT progress. On
top of this test approach, they built a recovery mechanism based on checkpoints.
The use of fault-tolerant message passing libraries specifically designed for MPSoCs
targeting fault reconfiguration are scarce. For instance, Kariniemi and Nurmi [KAR09] present a
similar fault-tolerant communication library targeting MPSoCs based on packet retransmission,
where timers and CRCs are used to detect packet losses and bit errors caused by transient faults
at the NoC links. In the case of a message lost, there is a retransmission mechanism triggered
based on acknowledgment (ACK) and negative acknowledgment (NACK). The NACK message is
generated after a timeout in the consumer, when the message is considered lost. Then the
consumer sends a NACK message, requesting the retransmission. The Authors do not discuss if a
faulty router is isolated or if it is circumvented in the case of a message considered lost. This may
indicate a case when a packet could be always routed to a faulty router and will never be received.
Furthermore, the Authors present the proposal, but without evaluating it.
Zhu and Qin [ZHU06] propose a fault-tolerant MPSoC executing a single program, multiple
data (SPMD) environment. In the SPMD approach, all PEs executes copies of the same program.
Then, if a tile is faulty, there is no instruction loss because another tile may execute the same
program. The proposed framework uses an MPI-like message-passing library. The Authors
evaluate a DSP application that has been replicated four times. This application was mapped in a
4x4 MPSoC, varying the number of non-faulty PEs, from 1 to 16. In the case of 16 available tiles,
no faulty tile exists in the configuration. The Authors observe that the fault-induced performance
degradation is smooth and predictable, with the exception of the configurations of very limited
number of non-faulty tiles. The reason is that, with three or more functioning titles, one of the tiles
is designated as the control tile while the computation tasks are mapped to the rest of the tiles.
Hébert et al. [HEB11] proposes a distributed fault-handling method for a NoC-based
MPSoC. Faults at the PE are detected with heartbeat messages exchanged among PEs. Each PE
has a watchdog timer capable of detecting missing or corrupted heartbeat messages. Besides,
fault detection, the method also comprises faulty PE isolation and recovering via task remapping.
As stated at the conclusion of [HEB11], the approach does not handle faults at the interconnect
network. The results showed silicon overhead of 26% of the overall logic composing the MPSoC
tile. The memory footprint overhead is about 12% when comparing to the baseline RTOS code.
Table 5 compares this review of prior work, showing that there are a few approaches
addressing one or two steps toward the design a dependable MPSoCs (fault mitigation, detection,
location, system reconfiguration, and system recovering). Approaches using a full MPI
implementation target cluster of computers, excepting [MAH08], which does not present the
implementation on a real platform. Most NoC-based systems adopt a partial MPI implementation
due to memory restrictions. Two similar approaches to this work [ZHU06][KAR09] only take into
account faults in the exchange of messages. As explained in the motivational example (Section
1.1), even retransmitting a message it will take the same faulty path, leading to an undelivered
message. Besides the FT communication protocol, an important contribution of this Chapter lies in
53
coupling the layered fault-tolerant routing method with the communication layer. At the network
layer, the proposed path recovery mechanism is capable of find fault-free paths with multiple NoC
faults. At the communication layer, we trigger the path reconfiguration when an undelivered
message is detected.
Table 5 – Comparison of different communication protocols adopting message passing.
Proposal
References
MMPI
[FU10]
MPI
Compatible?
YES simplified
SoC-MPI
[MAH08]
LA-MPI
MPI/FT
MMP
Zhu et al.
Proposed Work
NO
Communication
infrastructure
NoC
YES
NO
NoC/Bus
[AUL04]
[BAT04]
[KAR09]
YES
YES
YES simplified
YES
YES
YES
Cluster of Computers
Cluster of Computers
NoC
[ZHU06]
[WAC14a]
[WAC14b]
YES simplified
YES
NoC
YES simplified
YES
NoC
Fault Model
Broadly speaking, the three first steps (fault mitigation, detection, location) seem to be
mature, presenting different efficient test approaches for NoC-based systems. On the other hand,
reconfiguration and recovery steps for MPSoCs are still under development. We believe that welldefined protocol reconfiguration/recovery approaches are a necessary first step before addressing
the system-wide recovery (e.g. task migration and rollback-recovery) challenges. Thus, the
contribution of this Chapter lies on: (i) triggering of the path reconfiguration step, where an
alternative healthy path replaces the original faulty path; (ii) protocol recovery step, where the
protocol retrieves data lost during a faulty data transfer. At the application layer, the application
source code is not changed.
5.2
Proposed Fault-Tolerant Communication Protocol Description
When a given part of the NoC is faulty, eventually some packet will be not delivered. The
second contribution of this Thesis is to modify the original communication protocol to detect
these undelivered messages and to trigger the Seek Module. The required modifications to the
existing communicating protocol were:
•
All data packets are locally stored in the pipe before sending to the NoC. This
feature enables the packet retransmission since the source PE temporally keeps a
local copy of the packet;
•
For all delivered packets, an acknowledgment packet is transmitted from the target
to the source PE;
•
Each packet generated by a single PE receives a unique sequence number.
Figure 32 details the proposed fault-tolerant communication protocol. Different from the
original communication protocol presented in Figure 7 (page 25), the source PE does not release
the pipe when a message is requested. The slot with the message being transmitted assumes the
status “waiting acknowledgment” (label 1 in Figure 32). When the message is received, its
sequence number is verified. If it is the expected sequence number, the task can be scheduled to
run, the message is consumed (2), and the acknowledgment packet with the sequence number is
54
transmitted to the source PE (3). The last step of the protocol is to release the pipe slot, assigning
to its position an empty state (4).
TaskB
TaskA
Send()
pipe[i].status=USED
Requ
Msg
1
pipe[i].status=WAITING_ACK
4 pipe[i].status=EMPTY
Receive()
TaskStatus=WAITING
e st
MsgDeliv
ery
2
#1
TaskStatus=READY
ry#1
gDelive
s
M
k
c
A
3
Figure 32 – The proposed communication protocol to provide FT.
5.3
Faults in the Proposed FT Communication Protocol
The proposed FT communication protocol can be interrupted in four main situations. This
Section details each situation, explaining how the path search is triggered, and how the
communication continues using the new path.
5.3.1 Fault in the Message Delivery
This scenario, illustrated in Figure 33, shows the sequence chart of the protocol when there
is a fault in the path from A to B. Task B requests a message to Task A. Task A sends the
message delivery packet and waits for the acknowledgment packet. An Unresponsive
Communication Detection (UCD) scheme (described later in this Chapter) is used to detect that the
acknowledgment is not received. Then the source PE declares the target router (Task B) is
unreachable, and the fault-tolerant routing method is executed, and the faulty-free path stored in
the source’s local memory. After computing the new path, the packet is retransmitted to task B,
and the acknowledgment is received by task A.
TaskA
TaskB
Send()
MsgRequest
pipe[i].status=WAITING_ACK
MsgDeliv
Unresponsive
Communication
Detection
Receive()
TaskStatus=WAITING
ery#1
S eek
Backtrack
MsgDeliv
ery#1
RETRANSMISSION
AckMsgDelivery
Figure 33 – Protocol diagram of a message delivery fault.
5.3.2 Fault in Message Acknowledgment
In this scenario, illustrated in Figure 34, task B requests message from A, task A sends the
message to task B, and task B sends the acknowledgment back to task A. However, the
acknowledgment is not received due to a fault in the path from task B to task A. In this scenario,
task B successfully received the message, but the problem is that task A cannot release the pipe
slot having the consumed message, since it assumes that the message was not received.
55
TaskA
TaskB
Send()
MsgRequest
pipe[i].status=WAITING_ACK
Unresponsive
Communication
Detection
RETRANSMISSION
MsgDeliv
ery#1
AckMsgDelivery
Receive()
TaskStatus=WAITING
TaskStatus=READY
S eek
Backtrack
MsgDeliv
ery#1
Seek
Backtrack
Duetoduplicated
sequencenumber
AckMsgDelivery
Figure 34 – Protocol diagram of ack message fault.
Therefore, from this fault, the UCD scheme interrupts the PE holding task A, inducing the
process presented in the previous session (fault in the message delivery). As can be observed, the
search for the new path would not be needed, since the path AàB is faulty-free, but it is
impossible for task A to know why the acknowledgment was not received. This step corresponds to
the “retransmission” label in Figure 34.
Task B, in this case, receives the same packet with the same sequence number. Task B
discards this packet, signalizing an error in the acknowledgment path due a repeated sequence
number. Then PE holding task B starts the seek process to the PE holding task A, to find a faultyfree path for the acknowledgment packet. Once the new path to task A is received, the
acknowledgment packet is transmitted.
5.3.3 Fault in Message Request
Figure 35 shows the sequence chart of the protocol for a message request that was not
received (scenario similar to Figure 33). After the fault detection (UCD scheme), the seek process
is executed, and the last message request is retransmitted using the new path.
TaskA
TaskB
Send()
MsgRequest
pipe[i].status=WAITING_ACK
Receive()
TaskStatus=WAITING
MsgDeliv
ery#1
AckMsgDelivery
MsgRequest
Seek
Backtrack
Unresponsive
Communication
Detection
MsgRequest
Figure 35 – Protocol diagram of message request fault.
56
5.3.4 Fault While Receiving Packets
In wormhole packet switching networks, a packet might use several routers ports
simultaneously along the path from the source to the target router. If a faulty port is blocked while a
packet is crossing it, the packet is cut in two pieces as illustrated in Figure 36. The result of the
fault is two flows being transmitted inside the NoC: “F1” from task A to the faulty router; and “F2”
from the faulty router to the task B.
A
F1
F2
B
Figure 36 – Example of a faulty router in the middle of a message being delivered.
In fault mode, the wrapper logic around the faulty port signalizes that it can receive any flit
transmitted to it. The practical outcome is that all flits of flow F1 are virtually consumed by the faulty
port, avoiding fault propagation. This ensures that any data sent to a faulty port is immediately
dropped, and F1, in this case, disappears.
The flits belonging to the flow F2 reach the target PE. The target PE starts reading the
packet, and the kernel configures its DMA module to receive a message. The solution to discard
the incomplete packet was implemented in the NI. The number of clock cycles between flits is
computed. Here a fixed threshold was used, since the behavior of the NoC is predictable. If this
threshold is reached during the reception of a given packet, the NI signalizes an incomplete packet
reception to the kernel, and then it drops this packet.
5.4
Fault-Tolerant Communication Protocol Implementation
The proposed method targets permanent faults on the NoC. Assuming faults can block
some ports of the router, all packets sent to these ports are lost. To avoid the loss of these
packets, all packets supposed to use the faulty port must be rerouted. Therefore, we need to
explore the redundant paths of the network with the proposed reconfiguration and routing method
from the previous Chapter.
We also know that transient faults can lead to payload errors. These faults can be detected
by online error detection codes, such as CRC [LUC09], and are out of the scope of our work. On
the other hand, packets can be lost due to a misroute or by taking a faulty path. This chapter
focuses on the second kind of fault effect.
At the application level the task code responsible for calling the execution of Send() and
Receive() primitives does not require any modification. Therefore, the software designed for the
reference platform may be used with the new fault-tolerant protocol without modifications.
The implementation of the FT communication protocol comprises two parts:
•
Kernel: transfers the messages from the task memory space to the kernel memory
space and vice versa. This part implements the fault-tolerant communication
protocol and the deadlock avoidance algorithm described in Section 4.3;
57
•
Hardware: The hardware modules related to the communication protocol are the
network interfaces (NI) and the NoC routers.
The NoC received the path search module, presented in the previous Chapter. The idea is
that when an unresponsive communication is detected, the path search method is triggered. The
new path search is executed only once, when the unresponsive communication is detected. The
search returns a new path to the source PE. This path must be stored for future packet
transmissions, using source routing. In practice, each PE communicates with a limited amount of
PEs, enabling to use small tables for path storage. Thus, the path storage table can be
implemented either in hardware or software. If the path storage table is implemented in hardware,
the best module to place it is the NI, since it simplifies the kernel. For instance, the kernel injects a
packet to a given target in the NI. The NI checks the path table to verify if the target should be
transmitted using XY (i.e. the path is faulty-free) or source routing (if the original path has faults)
and, finally, the NI inserts the appropriate packet header.
The advantages to store the path in the NI include:
(i)
the network latency for packets sent with XY or source routing is the same, reducing
the performance penalties on faulty NoCs;
(ii)
the number of entries in the path table is not a function of the NoC size,
characterizing a scalable method;
(iii)
the process to find a new path is fast. Once the new path is computed, the
application’s protocol latency returns to its original performance.
Next sections describe the three detection schemes to detect unresponsive communication
developed during this Thesis. The first one, Adaptive Watchdog Timers, adopts a timeout to
deliver the message. If a message does not arrive in a parameterizable amount of time, the
message is set as lost, triggering the Seek Method. The second is named Auto-detection of Faulty
Paths, and uses the Seek Method together with a small module to detect messages that are routed
to faulty routers. After the message is detected, it sends a Drop Detection in the Seek Network to
notify the producer that the message has been lost. The third implementation simplified the FT
communication protocol, moving the acknowledgment to the next message request.
5.4.1 Adaptive Watchdog Timers [WAC14a]
The idea behind this approach is if any step of the communication protocol is stalled (e. g. a
task do not receive a message) for a given number of clock cycles, this message is considered
lost. Figure 37 represents six protocol transactions (in the x-axis) and their respective protocol
latencies (y-axis), assuming faulty and faulty-free transactions. These protocol latencies are used
to compute the AVG, the average protocol latency. K is a constant defined at design time per
application. Once the threshold of k*AVG is reached, the proposed method to find a new path is
fired and the packet is retransmitted using this new healthy path.
The AVG parameter is application dependent and computed at design time. The variability
of the protocol latency defines the k parameter. Dataflow applications require a small k value
because they a well-defined behavior. On the other hand, applications with tasks communicating
with several other tasks present higher protocol latency variability, requiring higher k values. The
user defines this parameter, k, based on an application profile.
time to find alternative
path (TFAP)
k*AVG
AVG
protocol latency
under fault (PLf)
fault free transaction
retransmission time
faulty transaction
...
protocol latency (K clock)
58
protocol transactions
Figure 37 – Comparing faulty and fault-free protocol latency.
Figure 37 also illustrates the main components of the protocol latency when a fault is
detected (PLf). It can be seen that most time is spent in k*AVG. Next, the proposed method is fired,
spending a small amount of time to find the alternative path (TFAP). Finally, this new path is used
for the packet retransmission. In addition, note in the last two protocol transactions that the
protocol latency returns to the nominal value. The latency returns to the nominal values because
the path calculation occurs only when the fault is detected, and the new path is used in all
subsequent packet transmissions. Equation 1 summarizes the components of PLf.
PLf=k*AVG+TFAP+retransmissiontime (1)
A fixed amount of time to start the seek process is not suitable for MPSoCs with
applications inserted at runtime. A communication intensive application will stall for a long period if
the fault timer threshold is too high. On the other side, a computation intensive application would
generate false fault alerts since these applications can take a long period between
communications. For these reasons, a new method is proposed, with an auto-detection of faulty
paths, described in the next section.
5.4.2 Auto-detection of faulty paths [WAC14b]
The previous implementation may introduce a large overhead in the application execution
time (AET). Section 5.5.1 shows that the use of watchdog timers may increase up to 68% the AET
compared to a faulty-free scenario.
The goal of the second FT communication protocol approach is to reduce the AET
overhead. The main idea relies on a hardware module responsible to notify the producer task that
a given packet was not delivered. The producer then starts the Seek Method and follows the same
steps as the previous approach.
The idea is to detect in the network when a packet is sent to a port that has been declared
faulty by a given test method. When it happens, this test method has to block the defective port (or
the entire router depending on the fault severity) by enabling the test wrapper presented in Chapter
3. When a neighbor router is about to send a packet to a faulty port, it fires a fault notification back
to the source PE. Then, the source PE can start the Seek Method to find a new path. Figure 38(a)
shows an example where a packet is sent from PE1 to PE2, but there is a faulty port in the middle
of the path, next to the PE3. Note that PE3 is aware of its neighbor’s faulty condition because of the
‘fault’ signals presented in Figure 38(b). PE3 sends the packet to the faulty neighbor, knowing that
this packet will be dropped and sends a fault notification back to the source PE (PE1).
59
5
5
5
SEEK
4
msgrequest
PE1
PE2
fault
TW
2
1
msgdelivery
PE3
fault
router
(a)
3
Drop Detection
router
PE3
(b)
Figure 38 – Auto-detection of faulty paths process.
The highlighted routers in Figure 38(b) present the hardware module responsible for firing
the Drop Detection message (the module is part of the router). PE3 is sending a packet to the faulty
router (label 1 in Figure 38 (b)). The Test Wrapper (label 2) discards this message and notifies the
Drop Detection module of PE3 that this connection is faulty (label 3). The Drop Detection module
contains an FSM and registers to store the source-target addresses of the discarded packet. When
a fault is detected, this module triggers the packet dropped process in the seek module (label 4).
Then, the seek network broadcasts the Drop Detection message to the neighbors (label 5),
addressed to the producer. When the producer receives a Drop Detection message, it sends a
seek clean to free the tables in the seek modules, then it triggers the new path discover procedure.
Figure 39 details the FT communication protocol with path reconfiguration and message
retransmission. The execution of Send() by a given task transfers the message to the OS level,
into the pipe structure. The following sequence of events occurs:
1. The task requesting data (Task B in PE2) injects a message request in the NoC.
2. The OS of PE1 answers the request, through a message delivery packet. In this Figure, the
message reaches a faulty router (fault in the west input port). The last faulty-free router,
PE3 (Figure 38 (b)), fires the fault recovery process.
3. The PE3 router discards all received flits at the faulty port and uses the seek network to
start a wave from PE3 (Drop Detection message) to PE1. This step is managed locally at
the router, does not require any action by the processor. Thus, the wave reaches PE1 in
few clock cycles.
4. When the FAULT wave reaches PE1, the processor is interrupted, and two actions are
managed by the OS: (i) the Seek Network launches a CLEAN wave to free the slots of the
seek memory used by the FAULT wave; (ii) the Seek Network launches a SEEK wave to
reach PE2.
5. The NoC is used to transfer the backtrack packet. At each hop, the packet receives the port
identification from which the SEEK wave entered.
6. When the backtrack packet reaches PE1, the Seek Network launches a CLEAN wave to
free the slots of the seek memory used by the SEEK wave, and the OS computes a
deadlock-free path. Using this example, the resulting path could be E0-E0-E0-N0-E0. This
path is stored, and all subsequent communications between PE1 and PE2 use this new path
based using source routing.
7. PE2 receives the message, schedules task B to execute and sends the acknowledgment
required by the protocol.
60
PE1
(c)
PE3
TaskA
executing
TaskA
executing
PE2
MsgReque
MsgDeliv
ery 2
n
etectio
DropD
st
TaskB
executing
1
faultdetected
3
4
clean
seek
TaskA
executing
5
OScomputea
deadlockfreepath
backtrack
clean
Msgdeliver
ywithso
TaskA
executing
6
urcerouting
TaskB
executing
AckMsgDelivery
7
Figure 39 – Sequence chart of protocol communication with Auto-detection of faulty paths.
5.4.3 Auto-detection of faulty paths with simplified protocol
The third approach simplifies the previous protocol, by removing the acknowledgment step.
One can see that in a communication between two tasks, the message acknowledgment is always
followed by a message request related to the next packet. For example, in Figure 40(1) Task B
acknowledges the reception of message number 1, then Task A frees the pipe in Figure 40(2). The
next step (Figure 40(3)) is the reception of a message request, where the second message status
is set as WAITING_ACK.
TaskB
TaskA
Send()
pipe[1].status=USED
pipe[1].status=WAITING_ACK
Send()
pipe[2].status=USED
2 pipe[1].status=EMPTY
pipe[2].status=WAITING_ACK
3
MsgRe
quest
MsgDeliv
ery#
1
1
1
very#
sgDeli
AckM
t
eques
MsgR
MsgDeliv
ery#
2
Receive()
TaskStatus=WAITING
TaskStatus=READY
TaskStatus=WAITING
TaskStatus=READY
Figure 40 – Example of acknowledgment message followed by a message request.
This third implementation merges the acknowledgment message with message request,
acknowledging the reception of a message and freeing the pipe slot. Figure 41 presents the new
FT communication protocol. This sequence chart shows the message exchanging of two
Send() – Receive() exchanges between Task A (tA) and Task B (tB).
1. tA executes a Send(), storing in the position one of the pipe the first message.
2. tB executes a Receive(), generating a message request to tA.
61
3. tA searches in all pipe positions for messages from tA to tB with WAITING_ACK status to
change its status to EMPTY. There is no message with this status, because it is the first
message to be sent.
4. Then, tA searches for the requested message from tA to tB with status equal to USED to
change its status to WAITING_ACK. This message (position one of the pipe) change the
status from USED to WAITING_ACK and is injected into the network.
5. tA executes a second Send(), storing the message in the second pipe slot.
6. tB receives the message, and sometime later, executes a new Receive(), generating a
second message request.
7. tA receives the message request and again searches in all pipe positions for messages
from tA to tB with WAITING_ACK status. One message fills this condition, and the pipe slot
number 1 is released, going to EMPTY state. As the message request has been received,
one can safely assume that the previous message has been received because the
communication has not generated a Drop Detection message.
8. tA sets the second message slot, with status USED, to WAITING_ACK, and inject into the
NoC the second message to tB.
9. The last step is to confirm the delivery of the last message of the task. The solution is to
verify if the task has already finished. In this case, the pipe slot can be set to EMPTY.
TaskB
TaskA
1
Send()
2
pipe[1].status=USED
3
Msg
e
Requ
st
Receive()
TaskStatus=WAITING
4
pipe[1].status=WAITING_ACK
5 Send()
MsgDeliv
ery
#1
6
pipe[2].status=USED
7 pipe[1].status=EMPTY
8 pipe[2].status=WAITING_ACK
9
Exit()
pipe[2].status=EMPTY
M
t
ques
sgRe
MsgDeliv
ery
#2
TaskStatus=READY
Receive()
TaskStatus=WAITING
TaskStatus=READY
Figure 41 – FT communication protocol without the acknowledgment message.
5.5
Results
This section first presents the experimental setup, the evaluation flow, and the applications
used as case study. Next, the results are divided into three sections, one for each method to detect
lost messages. The last sections present the area overhead and compare the proposed methods.
By using the isolation method presented in Chapter 3, it is possible to disable any individual
router link of the NoC. The router has five ports, but the local port is not used for fault injection, as
explained in Chapter 3. For this reason, faults can be injected in the eight remaining router links. A
small 3x3 network has 48 links (excluding the links in the chip’s boundary) in total. The number of
62
possible fault scenarios grows exponentially with the number of simultaneous faults. For instance,
1 fault requires C(48,1) = 48 scenarios and 2 faults requires C(48,2) = 1128 scenarios. It is not
possible to execute the evaluation manually since the number of fault simulations grows very fast.
For this reason, an automatic and parallel fault analysis flow, illustrated in Figure 42, has been
created [CAS14][PET12]. This flow is divided into five main steps.
Kernel Memory
(SystemC
initialized
Platform
memory)
Configuration App Object Code
(VHDL files) (VHDL initialized
memory)
Result Analysis
Phase
Fault Simulation
Phase
Log Analysis
n
tio
Scenario Generator
pseudo-random
fault scenario
generator
Fault
Database
Task1.c
Task2.c
TaskN.c
l i ca
MPSoC
configuration
p
Ap
MPSoC
Processor
FT Kernel
Description
(.C files) Description
(VHDL files)
(SystemC files)
Job submission
Fault Scenario Generation Phase
MPSoC Generation
scenario 1 – nok
scenario 2 – nok
...
scenario n – nok
Simulation Scenario
Figure 42 – Fault simulation flow.
The first phase, MPSoC Generation, is responsible for generating and compiling the
hardware and software of the MPSoC. As input, this phase receives the MPSoC hardware
description, the FT kernel, the application code and the MPSoC configuration file. The
configuration file contains parameters as MPSoC size, master processor location, local memory
size and abstraction level of hardware description. In our experiments, the processor and local
memory are described in cycle-accurate SystemC and the other modules are described in RTL
VHDL, enabling smaller simulation time when compared with the full VHDL description [PET12].
The MPSoC Generation phase generates the Simulation Scenario, which contains the compiled
kernel, the compiled application, and the MPSoC hardware model.
The fault scenario generation phase generates a database of faults scenarios to be
evaluated. These fault scenarios can be described manually, automatically, or a combination of
both. Manual scenarios are used to describe special fault scenarios with corner cases. Automatic
scenarios use pseudo-random generators to quickly generate hundreds of fault scenarios.
The fault simulation phase executes a fault simulation for each scenario in the database.
A grid computing resource distributes the simulation jobs in parallel among workstations. Each
simulation generates log files with the results, such as protocol latency, application execution time,
among other informations. These log files are parsed in the result analysis phase, extracting the
performance information and checking whether the application was able to execute with faults.
This phase also generates regression reports, charts, and tables used to compare each fault
scenario.
Four applications are used in this evaluation. Two of them are synthetic applications (called
basic and synthetic applications), used mainly for validation purposes, and two real applications
(with actual computation) that implements part of an MPEG encoder the DTW (dynamic time
warping) algorithm. These applications are decomposed in communicating tasks, illustrated in
Figure 43. This figure also shows the task mapping into the MPSoC.
63
M
A
A
B
start
B
print
idct
ivlc
iquant
iquant
print
M
start
(a) basic
M
D
A
C
E
C
P1
P1
P2
P7
bank
P3
P6
P5
P4
bank
recog
F
(c) synthetic
P2
recog
...
D
P8
F
E
B
M
idct
(b) mpeg
B
A
ivlc
P7
(d) DTW
Figure 43 – Evaluated applications, their task mappings and task graphs.
Table 6 summarizes the validation process for single fault scenarios. Without comparing the
approaches, we present some results that are common to the three methods:
•
•
•
•
scenarios: number of simulated scenarios;
executed scenarios (%): the percentage of possible scenarios. 100 means that all
possible fault scenarios were simulated;
affected scenarios: number of scenarios affected by faults. In this context, affected
means that at least one task fires a seek request;
node isolation: the number of fault scenarios that caused system failure due to a node
isolation.
Table 6 – Validation results with one fault.
basic synth MPEG DTW
scenarios
48
48
48
96
executed scenarios (%)
100
100
100
100
affected scenarios
8
12
8
91
node isolation
0
0
0
0
Results in Table 6 show that at least 8 fault injections affected the application execution
(affected scenarios row), and the proposed approach was able to find an alternative path, enabling
the application to finish its execution (node isolation row).
Table 7 details similar information, but assuming two faults per scenario. In this case, it is
not possible to simulate all possible fault scenarios since it would take several days of CPU time.
This way, 20% of the possible fault scenarios were randomly chosen for simulation.
Table 7 – Validation results with two faults.
basic synth MPEG DTW
scenarios
1213 1193
1213 2890
executed scenarios (%)
20
20
20
20
affected scenarios
215
394
251 1069
node isolation
2
3
3
10
64
The results in Table 7 show an increased number of affected fault scenarios (affectedscenarios row) but, for most of them, the proposed approach was able to determine an alternative
path. However, there were some system failures (node-isolation row). The reason is that tasks
might be isolated by faults, as illustrated in Figure 44. In this example, task A cannot send packets
to task B because both outgoing ports (east and south) are faulty. Redundant hardware or task
migration may restore the system execution when this scenario happens.
A
B
M
Figure 44 – Task isolation due to faults at east and south ports.
5.5.1 Performance Evaluation of Adaptive Watchdog Timers Implementation [WAC14a]
This section discusses how the faults might affect the Application Execution Time (AET)
using watchdog timers (Section 5.4.1). The AET includes computation and communication time,
and faults might affect the communication time by increasing the protocol latency (PLf). In the case
of watchdog timers, the PLf is related to k*AVG, the time to find an alternative path (TFAP), and the
time to retransmit the packet.
The TFAP includes the time required to execute the seek, backtrack, and path computation
steps, as presented in Section 4.4.2. The TFAP grows linearly with the number of hops between
the source to the target tasks. Each hop in the path requires in average 30 clock cycles for the
seek and the backtrack steps (15 clock cycles for each step). These steps are not time-consuming
because they are implemented in hardware. The path computation step takes about 200 clock
cycles per hop, and it is executed in software (kernel). For instance, the TFAP for a 10-hop path is
approximately 2,300 clock cycles, being 300 clock cycles consumed by seek and the backtrack
steps, and 2,000 clock cycles (200*10) to compute the path. The clock period in these experiments
is 10 ns. Thus, 2,300 clock cycles correspond to 23 us while the AET is in the order of
milliseconds.
The other relevant parameter for the AET is the time to fire the seek process, determined by
k*AVG. As discussed, the average protocol latency (AVG) is application-dependent, affecting in the
k value.
To evaluate the impact of k*AVG, the experiments presented in this section assumes a
fixed task mapping for each application (Figure 43), and a fixed fault scenario (i.e. fixed fault
location and fixed fault injection time). In this way, only the impact of k in the AET is evaluated. We
start with a k equal to 12, decreasing it until false seek request starts. Just a few fault scenarios are
presented in this section for the sake of clarity. In fact, although several scenarios inject fault at
different locations, they present the same impact in the AET.
Figure 45 presents the AET for the basic application considering four fault scenarios (labels
0 to 3), single fault per scenario, and k = {2, 4, 8, 12}. With the increase of k value, the AET under
fault condition increases up to 50%. Note that fault scenario also have a significant impact on the
variability of AET, meaning that the fault location and time, which are not under control of the
designer, have a relevant impact on AET in this method (watchdog timers). Scenario 0 (blue line)
represents the fault free AET used as reference. Note that the proposed method affects the fault-
65
free AET only when k = 2, because some false seek processes were fired, slightly increasing the
AET.
Figure 45 – Application execution time (AET) for basic application in scenarios with faults as a
function of parameter k. Each line represents a given scenario with one fault.
Figure 46 shows the AET for the MPEG application considering three fault scenarios (labels
0 to 2), single fault per scenario, and k = {2, 4, 8, 12}. The MPEG AET with faults is less than 11%
higher than the normal fault-free AET. One can observe that k has a smaller influence on AET
compared to the basic application. In addition, scenario 2 with k = 2 present a slight increase in
AET. This is caused by false seeks caused by the low k value, increasing the AET in the presence
of a fault. The scenario 0 (blue line) is the fault-free reference.
Figure 46 – Application execution time (AET) for MPEG application in scenarios with faults as a
function of parameter k. Each line represents a given scenario with one fault.
The simulation of scenarios with k< 8 in the previous experiments fired a large number of
false seeks. For this reason, Figure 47 shows the AET for the synthetic application considering six
fault scenarios (labels 0 to 5), single faults per scenario, and k = {8, 12}. The synthetic AET with
faults is about 13% higher than the normal fault-free AET. One can observe that k has a small
influence on AET. The fault scenario (fault location and fault injection time) has more impact than k
on the synthetic AET.
66
Figure 47 – Application execution time (AET) for synthetic application in scenarios with faults as a
function of parameter k. Each line represents a given scenario with one fault.
In conclusion, these experiments show that the variation of AET under fault situations may
be acceptable (~10%) when the computation time is higher than the communication time, as in real
benchmarks (MPEG example). Applications with more average latency (AVG) variability requires
larger k, therefore they might suffer larger AET variability under fault conditions.
The proposed method can only affect the AET under normal condition (fault-free) if k is too
small for the application, due to the generation of false seeks. A simple communication application
profile can help to determine the appropriate value of k.
5.5.2 Performance Evaluation of Auto-detection of faulty paths [WAC14b]
Table 8 presents the performance of the FT communication protocol for the scenario
presented in Figure 38, where a router has a single fault in the west input port, varying the distance
between the source processor (PE1) to the target processor (PE2). The adopted benchmark is the
MPEG, with 128-word messages (256-flit packets) transmitted from PE1 to PE2. There is no
congestion in the path since the goal is to characterize the protocol. In Table 8:
•
t0 – time required to inject a packet into the network, and the reception of the fault detection.
To inject a packet into the network, the OS: (i) configures the DMA module to transmit the
packet; (ii) restores the context of the current task; (iii) enables external interrupts to be
received. During this process, the fault detection interruption is received. As the time spent in
the NoC is smaller than the number of cycles consumed by the OS, this parcel of the protocol
is constant regardless the number of hops.
•
t1 – time spent between the fault detection by PE1 and the reception of the SEEK by PE2. The
fault detection packet interrupts the OS if PE1, triggers a clean wave to empty the fault
detection, and then triggers a seek to PE2. It increases, in average, 16 clock cycles per hop.
•
t2 – time to transmit the backtrack packet. Due to the hardware implementation, a small
number of clock cycles is observed, and it is a function of the number of hops.
•
t3 – corresponds to the number of cycles to compute the new path in the OS. It is also a
function of the number of hops.
•
t4 – time spent to deliver the packet using the faulty free path. The number of clock cycles is
constant for the same reason explained in the parcel t2àt3.
67
Table 8 – Number of clock cycles for each step of the FT communication protocol of Figure 39
varying the number of hops between PE1 and PE2.
PE1
(c)
PE3
TaskA
executing
TaskA
executing
st
MsgReque
MsgDeliv
ery 2
etectio
DropD
3
n
5
OScomputea
deadlockfreepath
t0
7
t0
769
769
769
769
769
769
t1
647
672
688
704
720
736
t2
48
51
63
75
87
99
t3
981
1084
1177
1291
1402
1499
t4
1023
1023
1023
1023
1023
1023
TOTAL
3468
3599
3720
3862
4001
4126
4
t1
t2
backtrack
clean
Msgdeliver
ywithso
TaskA
executing
1
faultdetected
clean
seek
TaskA
executing
hops #
2
3
4
5
6
PE2
6
t3
urcerouting
t4
7
AckMsgDelivery
In these scenarios, the amount of time to transmit a packet in a fault-free scenario is in
average 1670 clock cycles (Figure 48). Therefore, the worst-case overhead is observed in the 7hops scenario, corresponding to 2.46 times the time to transmit a message without fault. The
transmission of the first message (Figure 48) is faster because it is buffered while the consumer
task starts its execution. The third message requires path reconfiguration and message
retransmission. All others messages (4 to 8) are not penalized since the new path is also minimal,
and the time to transmit the message in both cases is the same. Consequently, the impact of the
fault-detection/path reconfiguration/retransmission in the application is small, since the process is
executed once per fault.
Figure 48 – Time spent to transmit eight 256-flit packets with and without fault using the proposed FT
communication protocol for the 8 first frames. The fault was detected in the third packet.
For the Auto-detection of faulty paths, three applications were evaluated with 1,200
scenarios. Two faults are injected at each scenario at the same moment, at random ports. The
simulations demonstrate that the method supports several simultaneous faulty ports, as long as the
number of faults does not physically split the application graph into disjoint graphs. Figure 49
presents the execution time for the 3,600 simulated scenarios.
68
Figure 49 – Applications execution time for each faulty scenario. Each dot in the figure represents a
given scenario.
Each application execution length is such that one message (out of 10) is affected by a
fault, ensuring a pessimistic evaluation scenario where the fault impact is not dissolved in an
extremely long application execution time. The worst-case execution time overhead was 2.53%,
1.87%, and 6.5% for the DTW, MPEG, and synthetic applications, respectively. The average-case
overhead was 0.17%, 0.09%, and 0.42%. Three reasons explain this remarkable result, compared
to the previous approach:
•
Independence of the fault location. The overhead is mostly induced by the moment
of the fault detection. If the PE is executing some OS function, the treatment of the
interruption due to the fault may be delayed. In other words, the time spent to treat
the interruption is much longer than the time to receive the fault notification.
•
Fast path reconfiguration. The retransmission of a given message roughly double
the packet transmission time (see Figure 48), with a small overhead due the fast
path reconfiguration (Table 8).
•
Overhead only in the undelivered message. After path reconfiguration, there is no
overhead in the communication since the new path is saved in the PE for future
transmission.
69
5.5.3 Performance Evaluation of Auto-detection of faulty paths with simplified protocol
The results for the Auto-detection of faulty paths with simplified protocol showed a decrease
of the AET compared to the previous approaches. The AET reduction comes from the
simplification of the communication protocol that removed the acknowledgement message. Results
concerning this implementation are detailed in the next section (Table 9).
5.5.4 Overall Evaluation of the FT Communication Protocols
Table 9 presents the AET worst-case for the three methods, with 1 and 2 simultaneous
faults. The DTW application was removed in the watchdog timer implementation. The reason is the
variability in the average latency, since multiple tasks communicate with the same task (e. g. recog
task in Figure 43), generating too many false seek requests.
Table 9 summarizes all the executed scenarios presenting the comparison for AET worstcase overhead for the three methods. The proposed methods were able to treat the injected faults.
The first approach showed the higher overhead due to the large delay to trigger the path search
method. The two other approaches showed a small impact in the AET when compared to
watchdog timer. Table 9 also shows that multiple simultaneous faults do not necessarily increase
the worst-case AET since the seek request associated to each fault is executed in parallel.
Table 9 – AET worst-case overhead for Watchdog timers and the two Auto-detection of Faulty Paths
approaches.
Approach
Watchdog Timers
Application
basic
mpeg
synthetic
DTW
number of fault(s)
1
2
67.46%
67.46%
68.47%
68.55%
35.99%
38.79%
N/A
N/A
Auto-detection of
Faulty Paths
number of fault(s)
1
2
N/A
N/A
2.16%
1.87%
7.5%
6.5%
3.46%
2.53%
Auto-detection of Faulty Paths
with simplified protocol
number of fault(s)
1
2
7.46%
5.85%
0.66%
0.43%
2.38%
1.45%
0.79%
0.60%
The hardware overhead due to the FT protocols comes from the drop_detection, with a total
cost 248/270 LUTs/FFs respectively (there is 8 of these modules in the router). The memory usage
overhead for the kernel with the proposed FT is 6.5 KB. For comparison, the memory usage
overhead in [HEB11] is 8.1 KB. Such values may be considered small for embedded systems.
5.6
Final Remarks
This Chapter presented a FT communication protocol for NoC-based MPSoCs. Both
hardware and software were integrated and validated on an existing MPSoC design described at
the RTL level. The proposed method was evaluated with synthetic and real applications while
permanent faults were injected into the NoC at runtime. The protocol automatically detects the
unreachable tasks and launches the search for a faulty-free path to target PE. This way, the
method enables applications to continue their execution, as long as there is at least a single
functional path to the target PE. The entire process of finding alternative paths takes typically less
than 2000 clock cycles or 20us. The protocol caused less than 1% in the total execution time of
computationally intensive applications in case of faults.
70
Note that the last modification, removing the acknowledgment message is valid only for the
communicating messages (message request and delivery). Therefore, other messages exchanged
with the MP, for example, should include acknowledgment messages to provide FT.
71
6 APPLICATION RECOVERY
Transport Layer
- FT Communication Protocol
- Fault Free Path Storage
- Path Search Triggering
Application Layer
- PE disabling
- Save task internal state
(checkpoints)
- Restore task (rollback)
5
6
A
PE
Network Layer
- Fault Tolerant Routing
- New Path Search
- Path Reconfiguration
4
B
Physical Layer
- Link Fault Detection
- Fault Isolation (Link
Wrapper)
3
This Chapter presents the fourth and last layer of the Thesis: the application layer. This
layer proposes a lightweight error recovery technique for multi-core systems.
This platform used to validate the methods presented in this Chapter is different from the
platform used in the previous layers. The implementation of this layer occurred during the Ph.D.
internship at the CEA-List laboratory, under the supervision of the Dr. Nicolas Ventroux. This
architecture, named P2012 [BEN12], is a multiprocessor platform, using shared memory as
communication model.
The goal of this layer is to execute a checkpoint method periodically, and then if a fault is
detected, the previously saved context is restored, allowing the system to continue its execution
with unaltered data.
6.1
The State-of-the-Art in Context Saving and Recovery Techniques
This section reviews the state of the in context saving and recovery methods. There are
common approaches targeting HPC and some for embedded systems.
6.1.1 ReVive [PRV02]
The Authors in [PRV02] present ReVive, a checkpoint/rollback mechanism for architectures
with processors, caches and memory interconnected by an off-chip network (e. g. HPC). They
implemented a partial separation technique with a logging checkpoint mechanism. This approach
proposes a partial separation, where checkpoint data and working data are one and the same,
except for those elements that have been modified since the last checkpoint.
The memory is modified in the following manner. When a write-back arrives, the method
checks whether this is the first modification of the line since the last checkpoint. If it is, the previous
content of the line is read from memory and saved in the log before the new content of the line is
written to memory. A modified line only needs to be logged once between a pair of checkpoints. To
this end, the directory controller is extended with one additional state bit for each memory line,
which the Authors call the Logged (L) bit.
This approach requires the memory to be divided into pages, with a hardware directory
controller responsible for the access to the memory.
72
For the evaluated applications, the approach imposes a worst-case area overhead of 2.5
MB for storing the checkpoint data. The results on a 16-processor system indicate that the average
error-free execution time overhead of using ReVive is only 6.3%.
6.1.2
Chip-level Redundant Threading [GON08]
In [GON08] the Authors propose a Chip-level Redundant Threading (CRT) to detect
transient faults on Chip Multiprocessors (CMPs). The approach is to execute two copies of a given
program on distinct cores and then compare the stored data. CRTR (CRT with Recovery) achieves
fault recovery by comparing the result of every instruction before commit. Once detecting different
results, the microprocessor could be recovered by re-executing from the wrong instruction. The
results showed that the performance overhead of the context saving when compared to the
baseline processor is approximately 30%.
6.1.3 Reli [LI12]
The Reli technique [LI12] proposes to change the micro-operations of instructions, which
stores registers and data memory. At runtime, Reli instructions execute not only native
functionalities (e.g., adding two operands of the ADD instruction), but also Reli functionality (e.g.,
generating checkpoint data of destination register for ADD instruction). They modify the microoperations in the instructions that change the state of the processor.
They adopted two stacks, named backup stacks, used for storing the registers in the
register file and for storing the data memory values that changed. For registers, the first time a
register is changed in a basic block, the old value of that register is stored in the backup stack. For
data memory, the old values are simply stored in the data memory stack along with the address of
the location.
After profiling (with SPEC INT 2006 and MiBench suites), the Authors concluded that the
maximum number of stack positions necessary was 66. Results show an overhead of 1.45% in
average execution time and 2.4% worst-case on a faulty-free scenario.
For fault injection, the Authors used a single bit-flip as the fault model and the number of
fault injections was 1000 for each application. They injected one fault for each application iteration.
For the evaluated applications, the worst-case recovery overhead is 62 clock cycles and the
highest application average recovery time is 17.9 clock cycles.
Regarding the cost of the proposal, the area overhead varies from 37.8% to 52.0% for the
evaluated applications (Reli implementations differ according to the application) compared to the
baseline implementation and 9.6% to 52.4% more leakage power.
6.1.4 DeSyRe [SOU13]
The DeSyRe project [SOU13] presents an MPSoC framework for FT purposes. The Authors
investigate techniques for checkpointing/migrating task state at each PE of the DeSyRe SoC. The
Authors claim that initial results indicate that maintaining duplicate copies across the NoC offers a
superior performance, but do not present this result. Furthermore, the duplicate copies have the
added benefit of protecting against NoC failures and disconnected nodes. According to these initial
results, the project assumes an architecture with a main memory being fault-free while the PEs and
local memories are fault-prone.
As error recovery technique, they propose the checkpoint and task re-execution. They
evaluated the MPSoC mapped to a Virtex6-based ML605 development board. The MPSoC
consists of a master Microblaze to host the runtime system, and 7 workers for tasks execution.
73
The evaluation is conducted with a matrix multiplication application with 24 tasks, varying
the matrix size. In addition, they consider three fault error rates: 4%, 20% and 41%. As this is an
ongoing work the Authors do not evaluate the checkpoint technique, therefore there is no results
related to the overhead in a fault-free scenario. However, the evaluation of the application reexecution in the scenario with 20% of tasks being faulty, the execution time doubles.
6.1.5 Barreto’s Approach [BAR15]
Barreto et al. [BAR15] propose an online fault recovery for embedded processors of
MPSoCs based on distributed memory. This approach automatically restarts affected applications
reallocating tasks to healthy processors. All steps are performed at the kernel level, without
changing user application code. The Authors employ HeMPS as the target platform.
The process is described as follows:
(1) The MP (Manager PE) receives a fault notification message, indicating a fault at a SP
(Slave PE).
(2) The MP constructs a set TF = {t1, t2, …, tn} representing the affected tasks running in the
defective SP.
(3) The MP identifies for each ti in TF the applications affected the fault, constructing a new
set TA = {A1, A2, …, Am}, being Ai an application to be frozen.
(4) The MP sends a freeze message for all tasks of each application in TA.
(5) The MP relocates only the tasks in TF to healthy SPs. All other unaffected tasks stay in
the same location. After the dynamic task mapping, the unaffected tasks receive the
new addresses for the relocated tasks.
(6) MP sends the relocation and the unfreeze messages for all tasks of each application in
TA.
The results shown the recovery overhead compared to the application execution time is
1,8% for one fault and for seven faults 5%. The recovery overhead is impacted only by the
reallocation time. In this proposal only the tasks affected by faults are reallocated, reducing the
overhead. Regarding the area overhead, there is no impact since all modifications were in the
software. There is no information for the kernel memory footprint overhead.
In fact, the Authors present a solution for application restarting. When a fault is detected in a
given PE, the application is restarted. Note that the application restarting is not a problem for
periodic applications, like video processing, representing the lost of few frames, in practice
imperceptible for the final user. However, the proposed approach cannot be applied for application
requiring the past state, as scientific computations.
6.1.6 Rusu’s Approach [RUS08]
The Authors in [RUS08] present a checkpointing method for NoC-based MPSoCs with
message passing. The proposal is a coordinated checkpoint, where a checkpoint initiator (a
dedicated task) asks to all tasks to perform checkpoints. As soon as possible after receiving this
request, a task takes a checkpoint, then sends an acknowledge message to the initiator. When the
initiator knows that all the tasks took their checkpoints, the new global checkpoint is validated and
the checkpointing procedure finishes.
As the initiator should send broadcast messages to all non-initiators, the Authors noticed
that it was inducing congestion and consequently increasing checkpointing latency. Then it is
74
proposed an optimization in the protocol (changing the complexity from O(n²) to O(n)) and in the
network with a more efficient broadcast method.
The simulations execute a single application with a unique task mapped to each PE. Each
task generates a uniform traffic with a constant rate of message injection for every task
(0.005/cycle) and a constant message length (64 bytes).
Results show that, with the modification in the protocol and in the broadcast, the overhead
of executing a checkpoint can be reduced to up to one order of magnitude in systems with more
than hundreds of PEs. The Authors evaluate the average checkpoint log size for each
implementation but do not discuss the size for a real application and how the dedicated memory to
the log can be estimated.
6.1.7 State of the art discussion
Table 10 presents a comparison of the evaluated works. Some of the methods shown an
overhead without faults smaller than 20%, considered by the Authors an acceptable overhead
[PRV02][LI12]. However, these approaches target distributed systems [PRV02] or require
modification in the ISA and dedicated hardware [LI12]. Basically all shared memory context saving
techniques are based on this two works.
In architectures with message passing, the largest problem is the overhead to synchronize
a global state for messages in each PEs [RUS08]. [BAR15] et al. adopt a simpler approach, since
the proposal do not present a global synchronization mechanism, restarting the whole application
in case of faults.
Other methods present a large area overhead and require redundant executions, wasting
processing resources [GON08]. In [SOU13], the Authors present an ongoing work, with no context
saving, just an application restarting approach.
Table 10 – Comparison of evaluated Fault-Tolerant Context Saving approaches.
[PRV02]
HPC
Communication
Model
shared memory
2.5MB memory
Fault free Performance
Overhead
6.3%
[GON08]
CMPs
shared memory
100%
30%
Reli
[LI12]
embedded singleprocessor
N/A
79.3% worst case
2.4% worst case/ 1.45%
avg
DeSyRe
[SOU13]
MPSoCs
N/A
N/A
N/A
Barreto’s
approach
Rusu’s
approach
[BAR15]
MPSoCs
message passing
zero (software
approach)
zero
5% worst case
[RUS08]
MPSoCs
message passing
N/A
N/A
N/A
shared memory
external - L3 to store
context
5.67%
17.33% - 28.34% (1 – 3 faults)
Proposal
References Target Architecture
ReVive
CRTR
Proposed Work
6.2
MPSoCs
Area Overhead
Recovery Overhead
59ms worst case/17ms avg
19% when compared to the fault free
execution
62 clock cycles worst case
17.9 clock cycles avg
doubles the execution time for 20% of
tasks being faulty
Fault-tolerant Reference Platform
The P2012 MPSoC architecture is an area- and power-efficient many-core architecture for
next-generation data-intensive embedded applications such as multi-modal sensor fusion, image
processing, or mobile augmented reality [BEN12][MEL12]. The P2012 contains multiple processor
clusters implemented with independent power and clock domains, enabling fine-grained power,
reliability and variability management. P2012 can reach 19 GOPS (with full floating point support)
in 3.8mm² of silicon with 0.5 W power consumption.
75
6.2.1 Architecture
Figure 50 presents an overview of the P2012 architecture. P2012 is a GALS fabric of tiles,
called clusters, connected through an asynchronous global NoC (GANOC) [THO10]. Each cluster
has access to an L2-shared memory and to an external L3-shared memory. Each P2012 cluster
aggregates a multi-core computing engine called ENCore and a cluster controller (CC).
The ENCore contains 16 STxP70-V4 processing elements (PEs). The STxP7 70-V4 has a
32-bit load/store architecture with a variable-length instruction-set encoding (16, 32 or 48-bit) for
minimizing the code footprint. Instructions can manipulate 32-bit, 16-bit or 8-bit data words, as well
as small vectors of 8 bits and 16 bits elements. The STxP70-V4 core is implemented with a 7stage pipeline for reaching 600MHz and it can execute up to two instructions per clock cycle (dual
issue).
Within the cluster, one STxP70-4 processor is implemented with a floating-point unit
extension (FPx). In this configuration, one single cluster can execute up to 16 floating point
operations per cycle. In addition, since all cores have independent instruction issue pipelines, there
is no single-instruction, multiple-data restriction on execution, which is a common restriction of
GPUs. This greatly simplifies application development and optimization.
Each core has an L1-16KB-private instruction cache. For the data cache, the PEs share an
L1-256KB tightly coupled data memory distributed in 32 banks (TCDM). This way, the PEs does
not have private data caches or memories, therefore avoiding memory coherency overhead.
Figure 50 – P2012 architecture.
The CC is the manager processor of its cluster. The CC encloses a processor named CCP
(Cluster Controller Processor), a DMA subsystem and two interfaces: one to the ENCore, one to
the GANOC. The CC processor also adopts the STxP70-V4 processor, 16-KB of program cache
and 32-KB of local data memory. The cluster controller processor, together with its peripherals, is
in charge of booting and initializing the ENCore. It also performs application deployment on the
ENCore. The DMA sub-system has two independent DMA channels. It performs the data block
transfers from the external memory to the internal memory and vice-versa while the various cores
are operating. The CC interconnects supports intra and inter-cluster communication.
6.2.2 Software Stack
The software stack is named HARS [LHU14] and it is based on a hardware-assisted
runtime software. It is composed of resource management features, multiple execution engines to
76
support different programming models, and synchronization primitives relying on a hardware
module named HardWare Synchronizer (HWS). ENCore provides scheduling and synchronization
acceleration by means of the HWS. The HWS includes a synchronization module that provides a
hardware-supported acceleration of various synchronization mechanisms: semaphores, mutexes,
barriers, joins, etc.
The HWS is dedicated to accelerate synchronization primitives on massively parallel
embedded architectures. It is designed as a peripheral to be integrated in architectures using
load/store operations, providing their runtime software with efficient synchronization
implementations even for architectures without atomic operations support. It can also remove
polling issues related to spin-lock operations. A specific set of software synchronization primitives
based on this hardware accelerator can be used by the different cores to perform synchronizations.
Thus, instead of using a software instruction requiring an atomic memory read/write access, the
synchronization primitives leverage the HWS atomic counters to implement locking. Moreover, the
runtime software uses sleep locks to put the processor in a waiting state until it is awakened when
the resource is free.
HARS proposes a small set of execution engines covering a wide range of parallel
programming styles. Two main execution engines are implemented: conventional multi-threading
for coarse grain parallel expression (suitable for thread-level or task-level parallelism) and
synchronous and asynchronous reactive tasks management for fine grain parallelism (suitable for
data-level parallelism).
Finally, an API enables the software designer to have access to all communication
primitives, parallel task execution triggering and control of the synchronization features presented
in other layers. The Software Stack is important as all the primitives of the saving and recovery
context are implemented in these layers.
6.2.3 Execution model
In the P2012, a conventional multi-threading execution model based on fork-join
mechanisms has been chosen. The PE that executes the fork is referred as master PE (PEm). Any
of the 16 PEs of the ENCore may be select as PEm. As showed in Figure 51, PEm executes the
sequential part of the application and can delegate tasks to other processors, parallelizing the
execution.
Figure 51 – Execution Model in P2012 with HARS. (1) master forks parallel tasks, (2) other PEs
execute the tasks, and (3) the master does the join.
PEs only executes tasks that were forked by PEm. To execute the fork, the PEm populates a
table T in a shared space with tasks to be executed. The fork procedure loads the local shared
77
memory within the cluster with the data and instructions to be executed by the parallel tasks. After
the load procedure is executed, each PE runs its task. When there are no more tasks to be
scheduled, the PEm waits until all tasks have finished their execution to join the tasks. Every PE
that is not doing a fork operation executes a scheduling loop. This loop searches for jobs to be
executed by scheduling ready tasks from table T.
Each PE accesses a dedicated shared memory space, which is released when the task
finishes its execution. At this point, the local memory in the cluster accessed by the PE has no
useful information for the execution on PEm and can be discarded. The fork-join process can be
repeatedly executed, but the PEm must wait all tasks to finish their job before the join.
6.3
Proposed Fault-Tolerant Method [WAC15]
As stated at the beginning of Chapter, this Thesis proposes a fault-tolerant approach to
tackle faults occurring in the processors. We assume that the fault detection scheme is out of the
scope of our approach. Implementations have already been proposed, as discussed in section 2.3.
All the modifications are implemented at the Software Stack (HARS). The simulation and
system evaluation is conducted with the model of the platform. This model is implemented at the
transaction level modeling (TLM).
According to the execution model, only the context of PEm is saved, as well as the global
shared memory space. Thus, the context saving/restoring process is performed before the fork and
after the join. This guarantees a coherent state for all PEs and eases the management of faults.
Thus, the execution context that must be considered is a structure composed by the 32
PEm’s registers, the .data section that stores all the shared uninitialized data, the .bss section that
stores all the shared initialized data, and the PEm stack which is locally stored in the CC L1-data
memory.
This context structure is stored in the L3 memory, accessible by all the clusters. All
accesses to this memory are made through the GANoC, inducing network traffic. The access time
is higher when compared to the local shared memory within the cluster. The context structure is
allocated at runtime according to the size the application needs.
For the results section, we modeled the access time for each memory. This time is taken
into account only for the first access. The subsequent operations are executed in one clock cycle.
Table 11 presents the access latencies.
Table 11 – Access time for memories in P2012.
Module
READ/WRITE IN L1
READ/WRITE ACCESS L2
READ/WRITE ACCESS L3
Access time (clock cycles)
1 CYCLES
21 CYCLES
200 CYCLES
The HARS software stack in P2012 does not allow the PEs to send an interruption to the
PEm. Then, it is not possible to interrupt the fork execution at the exact moment the fault is
detected. The proposal is to use an atomic counter faulty_PEs to store the information if the PE is
faulty or not. At the end of the parallel task execution, PEm verifies if there was an error in some
PE, reading its atomic counter. If a given PE is faulty, it is isolated from the execution processor list
and consequently will not execute any other task. Then, PEm starts the recover context procedure.
78
The fault insertion is executed by a dummy task that sets the faulty_PEs atomic counter if a
fault is detected. Then the management of isolation is performed by the use of atomic counter,
which is checked each time the scheduler function on each PE is called.
Figure 52 presents the proposed Fault-Tolerant execution model. Before the sequential
execution is forked, the master saves the application context. This means that it stores in the
global shared memory (in this order): (1) all processor registers; (2) its stack: (3) its .bss section
and (4) its .data section. At the end of the fork/join process, the master checks if any of the PEs
detected a fault and if needed, it triggers the recover context procedure. If there was no fault, the
execution continues normally.
Figure 52 – Fault-Tolerant Execution Model: In (1) the master executes a context saving and in (3) it
verify if there was a fault, if positive, the context is restored and the fork is re-executed avoiding the
faulty PE.
6.4
Evaluation of the Proposed Fault-Tolerant Method
This section evaluates the overhead induced by the FT proposal. All scenarios are executed
in the P2012’s SDK released by ST Microelectronics. For the results, only 1 cluster is considered,
and all the communications between tasks are made through the global shared memory space in
L3 and memory accesses are made through GANOC.
Two applications are adopted as benchmarks. The first application is synthetic, with it task
graph presented in Figure 53. A parameterizable number of NOP instructions (N) defines the task
size. It is also possible to parameterize the number of tasks (T), and the number of iterations (R).
Figure 53 – Task Graph of the synthetic application. (1) The PEm executes the context saving, (2) the
fork splits the execution in T tasks, each one executing N number of NOP instructions, and (3) this
process is replicated R times.
79
The second application is an industrial application named Human Body Detection and
Counting (HBDC). It consists of processing an image sequence to determine the background
image and subsequently the moving objects of the scene. The first phase uses the Mixture of
Gaussian (MoG) technique [MCL00]. It is forked in 60 tasks, each one taking around 340,000 clock
cycles to execute. Then, the remaining tasks are sequential. The moving objects are classified to
determine whether they correspond to human shapes where 64 image frames are processed.
6.4.1 Evaluation of the method with Synthetic Application
Figure 54 presents the evaluation of the impact of the context saving varying the task size
(N). The time to save the context is not a function of N. Thus, the size of the parallelized tasks
should mask the context saving overhead. As shown in the Figure, the execution overhead
reduces as N increases. A task with 10,000 NOPs has an overhead close to 20%, which is
considered an acceptable overhead. The next experiments use N=10,000 as reference for the task
size.
Figure 54 – Execution time overhead varying the number of NOPs in each task (T=10, R=10).
The next experiment evaluates the impact of the context saving, varying the size of the
sections .data and .bss (Figure 55). The amount of data to save is the main limitation of the
approach. A trade-off has to be defined between the tasks’ execution time and the context data
size. Then the programmer can choose an acceptable overhead cost of the context saving.
Figure 55 – Execution time overhead of context saving changing the context data size from 10 to 10k
words of 32 bits (T=10, R=10, N=10,000).
80
The context saving is disabled to enable the evaluation of the fork overhead. Figure 56
shows the execution time overhead varying R. The Figures shows that the fork execution takes
approximately 6% of the execution time. Since this overhead is independent of the number of
repetitions, the fork/join process has a limited impact on the performance.
Figure 56 – Fork overhead varying the number of repetitions (T=10, N=10,000).
Figure 57 shows the execution time for the synthetic application without context saving
(noFT), with the FT method and no injected fault (0faults), and with the FT method and a variable
number of injected faults. The context saving implies a 35% overhead (N=10,000), and for each
injected fault there is an increase of 15% for the context restoring and fork rescheduling.
Figure 57 – Application execution time overhead for scenarios with no context saving, and the
overhead for scenarios where there is overhead increasing the number of faults (T=10, R=10,
N=10,000).
6.4.2 Evaluation of the method with HBDC Application
Figure 58 shows the execution time overhead when context saving is executed according to
a variable number of frames. Saving the context at each eight frames increases the execution time
81
by 5,67%. This means that the background images will be restored as it was eight frames back if a
fault is detected. For this application, the checkpointing frequency has only a QoS impact that will
depend on the application frame rate. With a high frame rate, losing some frames will not affect the
application, resulting in a good tradeoff between performance and quality.
Figure 58 – Execution time overhead without faults when executing context saving from each frame
to each 16 frames. The bars show the context saving overhead and the execution time.
Figure 59 presents five executions of the HBDC application, assuming context saving at
each eight frames. In the first column (noFT), there is no context saving, being the baseline
execution time. The second column shows the overhead induced by the context saving with no
fault insertion (5.67% compared to baseline). The last three columns show the overhead for one,
two and three faults in frames of different saved contexts of the application. Note that the
percentage represents the overhead compared to the baseline and the highlighted part represents
only the context restoring overhead. As there are tasks to be re-executed, the task execution time
increases when the number of restored context grows.
Figure 59 – Application execution time with no context saving, the overhead induced by the context
saving and the overhead induced by the context saving plus the recovery time for one, two and three
faults. The percentages represent the overhead compared to the baseline. The highlighted part
represents the time executed restoring the context.
82
6.5
Final Remarks
This Chapter presented a Fault-Tolerant Context Saving for a state-of-the-art shared
memory MPSoC. Results show that the proposal was validated and can recover applications from
faults occurring in PEs. Execution with an industrial application shows a good tradeoff between
execution time overhead with no faults (5.67%) and with faults (17.33% - 28.34%). The proposal
does not imply in hardware overhead or redundant executions, as other works in the state-of-theart.
The proposal presented a method for shared memory architectures, adapting existing
techniques for distributed systems in embedded systems. The original contributions rely on (i) in
the evaluation of the cost of the techniques in the P2012 architecture and (ii) an isolation technique
to isolate a faulty internal core.
We adopt two assumptions in the current Chapter: (i) fault detection is out of the scope of
this research; (ii) there are no pragmas or code added by the software designer, allowing context
saving at any moment of the application execution.
A limitation of the proposed method is that it is very dependent to the communication model,
shared memory. The speedup of the overall application is limited to the part where the application
can be parallelized. Thus, our limitation is coupled to the fork-join approach.
The first step towards fault-tolerance for distributed memory architectures was presented in
[BAR15]. A careful evaluation of context saving/rollback is an important work to be accomplished in
a short term for the HeMPS platform. As stated in the state-of-the-art, this kind of communication
model presents a synchronization problem, being necessary to have a centralized node to
coordinate all the tasks to take their checkpoints.
83
7 CONCLUSION
From one side, MPSoCs provide a massive processing power offered by the large number
of PEs. From the other side, the failure probability due to the technology scaling also increases.
Runtime adaptation is a key to providing reliability of future SoCs. Employing only design time
techniques, such as TMR, designs would have to be more and more conservative because they
would spend more silicon area on the chip to provide some infrastructure to support faults. On the
other hand, runtime techniques provide a graceful degradation of the system. Runtime techniques
reconfigure the system to continue delivering their processing services, despite defective
components due to the presence of permanent and/or transient faults throughout the system
lifetime.
The Introduction of the Thesis stated the following hypothesis: “A layered approach can
cope with faults in the communication infrastructure and in the processing elements in such a way
to ensure a correct operation of the MPSoC, even in the presence of multiple faults. The proposed
methods increase the lifetime of current MPSoCs due to the isolation of defective routers, links,
and processing elements. The remaining healthy components continue to execute the applications
correctly.”
The layered approach enabled to divide the fault tolerance problem in independent parts.
Each layer is responsible for solving part of the problem, as an upper layer uses the features of the
lower layer to solve another part of the problem. For example, the network layer is responsible for
finding a new path when a fault is detected at some router in the path. The upper layer (transport)
is responsible for triggering the network layer, ensuring the correct communication between
processing elements. Consequently, the first part of the above hypothesis is proven.
Results demonstrated that even with several defective routers or processing elements the
MPSoC continued to execute the applications correctly, therefore, increasing the system lifetime.
Such results confirm the second part of the hypothesis.
The method has original contributions in all layers, being the first proposal to addresses a
comprehensive fault-tolerant approach:
•
In the physical layer, we propose the modification of test wrapper cells for FT purposes. With
this modification, it is possible to isolate the PE and the router that is faulty, avoiding Byzantine
faults. This modification can be coupled with a given test scheme to trigger the FT layered
approach. In addition, this modification enabled the design of a CRC fault detection scheme in
the HeMPS platform in the scope of our research group [FOC15], which is not part of the
contribution of the Thesis.
•
At the network layer, we propose a fast path finding method, being able to find faulty-free
paths in less than 40 microseconds for a worst-case scenario.
•
In the transport layer, different approaches were evaluated being capable of detecting a lost
message and start the message retransmission. The results show that the overhead to
retransmit the message is 2.46X compared to the time to transmit a message without fault,
being all other messages transmitted with no overhead.
•
At the application layer, the entire fault recovery protocol executes fast, with a low execution
time overhead with (5.67%) and with faults (17.33% - 28.34%). Table 12 resumes each layer
contribution, comparing it with the OSI model.
84
Table 12 – Simplified view of the OSI model with the added features for fault tolerance in each layer.
Layer
Architectural Features
Fault tolerance features
Physical
(Chapter 3)
Duplicated physical channels per link
(16-bit flit)
CRC added to data links and CRC in routers.
(in the scope of the research group [FOC15])
Synchronous credit-based flow control
Test wrappers; packet discarding
Adaptive Routing
Auxiliary NoC; degraded mode operation; faulty-free path search
Data Link
(Chapter 3)
Network
(Chapter 4)
Transport
(Chapter 5)
Application
(Chapter 6)
7.1
Message passing (OS level);
Communication API
Applications with checkpoints
Fault-tolerant communication protocol; packet retransmission
Task isolating and application checkpoint/rollback
Limitations of the Current Proposal
The proposal has limitations, which are reference for future works:
•
At the physical layer, the proposal requires a NoC with duplicated physical channels, in such a
way to enable full adaptive routing.
•
The Manager Processor (MP) is assumed fault-free. The FT communication method presented
in Chapter 5 assumes the communication between PEs. Therefore, the communication
between the MP and the PE is prone to faults that can lead to a system failure. The FT
communication protocol should be extended to cope with faults in the path to the MP, and it
should be possible to remap the MP into another processing element.
•
The external memory is assumed fault-free, and the interface with this memory is a single
point of failure. Replicating the external memory and the PEs with access to it could mitigate
this limitation. As a second advantage, it will also remove the bottleneck to map all tasks in the
system initialization.
•
At the network layer, it is necessary a processor to execute the path computation. A router
connected to a “passive” IP, as a memory bank, is not able to compute the deadlock-free path.
A possible solution is to use the MP to handle the path computation for such IPs.
•
The FT communication protocol was designed for message-passing communication. Shared
memory communication requires adaptation in the proposed protocol.
7.2
Future Works
After the development of the seek network, the simulated scenarios shown the underuse of
the seek network, even with multiple faults. Thus, the seek network could be used for other
purposes, as a monitor for aging effects, path congestion and to transmit control packets.
The test recovery method can act as an aging monitor, being managed by the MP. If a
given channel is affected by a large number of transient faults, this may signalize a wear out in the
channel, and the test module may notify the MP to disable this port permanently. Note that this
aging monitor is not limited to the channels of the router, but also can be used for the PEs.
85
A future implementation is to expand the implementation in [BAR15] to execute a
Checkpoint-Recovery method for message passing. The proposal can be a mix of the approach
proposed in this Thesis [WAC15] with the proposal of synchronization in [RUS08].
Another future work includes the use of more than one task repository, reducing the
bottleneck to map tasks and removing the single point of failure, as discussed in the previous
session.
86
REFERENCES
[AGA09]
Agarwal, A.; Iskander, C.; Shankar, R. “Survey of Network on Chip (NoC) Architectures
& Contributions”. Journal of Engineering, Computing and Architecture, vol. 2-1, 2009.
[ALA03]
Alam, M.A. “A critical examination of the mechanics of dynamic NBTI for PMOSFETs”.
In: Electron Devices Meeting, 2003, 4p.
[ALH13]
Alhussien, A.; Verbeek, F.; van Gastel, B.; Bagherzadeh, N.; Schmaltz, J. “Fully
Reliable Dynamic Routing Logic for a Fault-Tolerant NoC Architecture”. Journal of
Integrated Circuits and Systems, vol. 80-1, March 2013, pp. 43-53.
[ASA06]
Asanovic, K.; Bodik, R.; Catanzaro, B.; Gebis, J.; Husbands, P.; Keutzer, K.; Patterson,
D.; Plishker, W.; Shalf, J.; Williams, S.; Yelick, K. “The Landscape of Parallel
Computing Research: A View from Berkeley”. Technical Report, Electrical Engineering
and Computer Sciences – University of California at Berkeley, 2006
[AUL04]
Aulwes, R.T.; Daniel, D.J.; Desai, N.N.; Graham, R.L.; Risinger, L.D.; Taylor, Mark A.;
Woodall, T.S.; Sukalski, M.W. “Architecture of LA-MPI, a network-fault-tolerant MPI”.
In: Parallel and Distributed Processing Symposium, 2004, 10p.
[BAR15]
Barreto, F.; Amory, A.; Moraes, F. “Fault Recovery Protocol for Distributed Memory
MPSoCs”. In: ISCAS, 2015, pp. 421-424.
[BAT04]
Batchu, R.; Dandass, Y. S.; Skjellum, A.; Beddhu, M. “MPI/FT: A Model-Based
Approach to Low-Overhead Fault Tolerant Message-Passing Middleware”. Journal
Cluster Computing, vol. 7-4, October 2004, pp. 303-315.
[BEN02]
Benini, L.; De Micheli, G. “Networks on chips: a new SoC paradigm”. Computer, vol.
35-1, January 2002, pp. 70-78.
[BEN12]
Benini, L.; Flamand, E.; Fuin, D.; Melpignano, D. “P2012: Building an ecosystem for a
scalable, modular and high-efficiency embedded computing accelerator”. In: DATE,
2012, pp. 983–987.
[BOR07]
Borkar, S. “Thousand core chips: a technology perspective”. In: DATE, 2007, pp. 746749.
[BOY93]
Boyan, J.; Littman, M. “Packet Routing in Dynamically Changing Networks: a
Reinforcement Learning Approach”. In: Advances in Neural Information Processing
Systems, 1993, pp. 671-678.
[BUS05]
M. L. Bushnel and W. D. Agrawal, “Essentials of Electronic Testing for Digital, Memory,
and Mixed-Signal VLSI Circuits”. Boston: Springer, 2005, 690p.
[CAR08]
Carara, E.; Oliveira, R. P.; Calazans, N. L. V.. “A new router architecture for HighPerformance intrachip networks”. Journal Integrated Circuits and Systems, vol. 3-1, pp.
23-31, 2008.
[CAR09]
Carara, E. A.; Oliveira, R. P.; Calazans, N. L V; Moraes, F. G. “HeMPS - a framework
for NoC-based MPSoC generation”. In: In: ISCAS, 2009, p. 1345-1348.
[CAR10]
Carara, E.; Moraes, F. “Flow Oriented Routing for NOCS”. In: SOCC, 2010, pp. 367370.
[CAS14]
Castilhos, G.; Wachter, E.W.; Madalozzo, G.A.; Erichsen, A.; Monteiro, T.; Moraes,
F.G. “A Framework for MPSoC Generation and Distributed Applications Evaluation”. In:
ISQED, 2014. 4p.
[CHI11]
Wang, C.; Hu, W.; Lee, S.; Bagherzadeh, N. “Area and power-efficient innovative
congestion-aware Network-on-Chip architecture”. Journal of Systems Architecture, vol.
57-1, January 2011, pp. 24-38.
87
[CLE13]
CLEARSPEED.
“CSX700”.
Available
http://www.clearspeed.com/products/csx700.php. December 2013.
at:
[CON09]
Concatto, C.; Matos, D.; Carro, L.; Kastensmidt, F.; Susin, A.; Cota, E.; Kreutz, M.
“Fault tolerant mechanism to improve yield in NoCs using a reconfigurable router”. In:
SBCCI, 2009, 6p.
[COP08]
Coppola, M.; Grammatikakis, M.; Locatelli
R.; Maruccia, G.; Pieralisi L. “Design of
Cost-Efficient Interconnect Processing Units: Spidergon STNoC”. CRC Press, 2008,
288 p.
[COT12]
Cota, É.; Amory, A. M.; Lubaszewski, M. S. “Reliability, Availability and Serviceability of
Networks-on-Chip”. Springer, 2012, 209p.
[DEO12]
DeOrio, A.; Fick, D.; Bertacco, V.; Sylvester, D; Blaauw, D.; Jin Hu; Chen, G. “A
Reliable Routing Architecture and Algorithm for NoCs”. IEEE Transactions on CAD of
Integrated Circuits and Systems, vol. 31-5, 2012, pp. 726-739.
[DOO14]
Doowon Lee; Parikh, R.; Bertacco, V. “Brisk and limited-impact NoC routing
reconfiguration”. In: DATE, 2014, pp. 1-6.
[EBR13]
Ebrahimi, M.; Daneshtalab, M.; Plosila, J.; Tenhunen, H., “Minimal-path fault-tolerant
approach using connection-retaining structure in Networks-on-Chip”. In: NOCS, 2013,
4p.
[FEE13]
Feehrer, J.; Jairath, S.; Loewenstein, P.; Sivaramakrishnan, R.; Smentek, D.; Turullols,
S.; Vahidsafa, A. “The Oracle Sparc T5 16-Core Processor Scales to Eight Sockets”.
IEEE Micro, vol. 33-2, Mar-Apr 2013, pp. 48-57.
[FEN10]
Feng, C.; Lu, Z.; Jantsch, A.; Li, J.; Zhang, M. “A Reconfigurable Fault-Tolerant
Deflection Routing Algorithm Based on Reinforcement Learning for Network-on-Chip”.
In: NoCArc, 2010, pp. 11-16.
[FIC09]
Fick, D.; DeOrio, A.; Chen, G.; Bertacco, V.; Sylvester, D; Blaauw, D. “A Highly
Resilient Routing Algorithm for Fault-Tolerant NoCs”. In: DATE, 2009, pp. 21-26.
[FLI07]
Flich, J.; Mejia, A.; Lopez, P.; Duato, J. “Region-Based Routing: An Efficient Routing
Mechanism to Tackle Unreliable Hardware in Network on Chips”. In: NOCS, 2007, pp.
183-194.
[FLI12]
Flich, J.; Skeie, T.; Mejia, A.; Lysne, O.; Lopez, P.; Robles, A.; Duato, J.; Koibuchi, M.;
Rokicki, T.; Sancho, J.C. “A Survey and Evaluation of Topology-Agnostic Deterministic
Routing Algorithms”. IEEE Transactions on Parallel and Distributed Systems, vol. 23-3,
March 2012, pp.405-425.
[FOC15]
Fochi, V.; Wachter, E.; Erichsen, A.; Amory, A.; Moraes, F. “An Integrated Method for
Implementing Online Fault Detection in NoC-Based MPSoCs”. In: ISCAS, 2015, 15621565.
[FU10]
Fu, F.; Sun, S.; Hu, X.; Song, J.; Wang, J.; Yu, M. “MMPI: A flexible and efficient
multiprocessor message passing interface for NoC-based MPSoC”. In: SOCC, 2010,
pp. 359-362.
[GAR13]
Garibotti, R.; Ost, L.; Busseuil, R.; Kourouma, M.; Adeniyi-Jones, C.; Sassatelli, G.;
Robert, M. “Simultaneous multithreading support in embedded distributed memory
MPSoCs”. In: DAC, 2013, 7p.
[GE00]
Ge-Ming C. “The odd-even turn model for adaptive routing”. IEEE Transactions on
Parallel and Distributed Systems, vol.11-7, July 2000, pp. 729-738.
[GHA82]
Ghate, P.B. “Electromigration-Induced Failures in VLSI Interconnects”. In: Reliability
Physics Symposium, 1982, 8p.
[GIZ11]
Gizopoulos, D.; Psarakis, M.; Adve, S.V.; Ramachandran, P.; Hari, S.K.S.; Sorin, D.;
Meixner, A.; Biswas, A.; Vera, X. “Architectures for online error detection and recovery
in multicore processors”. In: DATE, 2011, 6p.
88
[GON08]
Gong, R.; Dai, K.; Wang, Z. “Transient Fault Recovery on Chip Multiprocessor based
on Dual Core Redundancy and Context Saving”. In: ICYCS, 2008, pp. 148–153.
[HEB11]
Hebert, N.; Almeida G. M.; Benoit, P.; Sassatelli G.; Torres, L. “Evaluation of a
Distributed Fault Handler Method for MPSoC”. In: ISCAS, 2011, pp. 2329-2332.
[HEN13]
Henkel, J.; Bauer, L.; Dutt, N.; Gupta, P.; Nassif, S.; Shafique, M.; Tahoori, M.; Wehn,
N. “Reliable on-chip systems in the nano-era: Lessons learnt and future trends”. In:
DAC, 2013, 10p.
[IBM13]
IBM. “The Cell Project”. Available at: https://www.research.ibm.com/cell/cell_chip.html.
December 2013
[INT13a]
INTEL.
“Teraflops
Research
Chips”.
Available
http://www.intel.com/pressroom/kits/Teraflops/index.htm. December 2013.
at:
[INT13b]
INTEL,
“Single-Chip
Cloud
Computer:
Project”.
Available
http://www.intel.com/content/www/us/en/research/intel-labs-single-chip-cloudcomputer.html. December 2013
at:
[ITR15]
International Technology Roadmap for Semiconductors. “ITRS 2011 edition”. Accessed
in: http://www.itrs.net/reports.html. April 2015.
[JAV14]
Javaid, H.; Parameswaran, S.
Multimedia”. Springer, 169p. 2014.
[JER05]
Jerraya, A. A.; Wolf, W. “Multiprocessor Systems-on-Chips”. Morgan Kaufmann, 2005,
602p.
[KAR09]
Kariniemi, H.; Nurmi, J. “Fault-Tolerant Communication over Micronmesh NOC with
Micron Message-Passing Protocol”. In: SOC, 2009, pp. 5–12.
[KER14]
Kerkhoff, H.G.; et al. “Linking aging measurements of health-monitors and
specifications for multi-processor SoCs”. In: DTIS, 2014. pp. 1-6.
[KOI08]
Koibuchi, M.; Matsutani, H.; Amano, H.; Pinkston, T. M. “A Lightweight Fault-Tolerant
Mechanism for Network-on-Chip”. In: NOCS, 2008, 10p.
[LEE61]
Lee, C. Y. “An Algorithm for Path Connections and Its Applications”. Transactions on
Electronic Computers, September 1961, pp. 346-365.
[LHU14]
Lhuillier, Y.; Ojail, M.; Guerre, A.; Philippe, J.; Chehida, K.; Thabet, F.; Andriamisaina,
C.; Jaber, C.; David, R. “HARS: A hardware-assisted runtime software for embedded
many-core architectures”. ACM Transactions on Embedded Computing, 2014, 25 p.
[LI12]
Li, T.; Ragel, R.; Parameswaran, S. “Reli: hardware/software checkpoint and recovery
scheme for embedded processors”. In: DATE, 2012, pp 875–880.
[LIN91]
Linder, D.H.; Harden, J.C. “An Adaptive and Fault Tolerant Wormhole Routing Strategy
for k-ary n-cubes”. Transactions on Computers, vol. 40-1, 1991, pp. 2-12.
[LUC09]
Lucas, A.; Moraes, F. “Crosstalk fault tolerant NoC - design and evaluation”. In: IFIP
VLSI-SOC, 2009, 6 p.
[MAH08]
Mahr, P.; Lorchner, C.; Ishebabi, H.; Bobda, C. “SoC-MPI: A Flexible Message Passing
Library for Multiprocessor Systems-on-Chips”. In: International Conference on
Reconfigurable Computing and FPGAs, 2008, pp. 187-192.
[MAR09]
Marinissen, E.J.; Zorian, Y., “IEEE Std 1500 Enables Modular SoC Testing”. Design &
Test of Computers, vol. 26-1, Jan-Feb 2009, pp. 8-17.
[MAR13]
Maricau, E.; Gielen, G. “Analog IC Reliability in Nanometer CMOS”. Springer, 2013.
[MAT13]
Matos, D.; Concatto, C.; Kologeski, A.; Carro, L.; Kastensmidt, F.; Susin, A.; Kreutz, M.
“A NOC closed-loop performance monitor and adapter”. Journal Microprocessors &
Microsystems, vol. 37-6, August 2013, pp. 661-671.
“Pipelined
Multiprocessor
System-on-Chip
for
89
[MCL00]
McLachlan, G. J.; Peel, D. “Finite mixture models”. New York: Wiley, 2000.
[MEL12]
Melpignano, D.; Benini, L.; Flamand, E.; Jego, B.; Lepley, T.; Haugou, G.; Clermidy, F.;
Dutoit, D. “Platform 2012, a many-core computing accelerator for embedded SoCs:
Performance evaluation of visual analytics applications”. In: Design Automation
Conference (DAC), 2012, pp.1137-1142.
[NVI13]
NVIDIA, “NVIDIA KEPLER GK110 Next-Generation CUDA Compute Architecture”.
Available at: http://www.nvidia.com/object/nvidia-kepler.html. December 2013.
[PET12]
Petry, C.; Wachter, E.W.; Castilhos, G.; Moraes, F.G.; Calazans, N. “A Spectrum of
MPSoC Models for Design Space Exploration and Its Use”. In: RSP, 2012, pp. 30-35.
[PRV02]
Prvulovic, M.; Zheng Zhang; Torrellas, J. “ReVive: cost-effective architectural support
for rollback recovery in shared-memory multiprocessors”. In: ISCA, 2002, pp. 111–122.
[PSA10]
Psarakis, M.; et al. “Microprocessors software-based self-testing,” IEEE Design and
Test of Computers, v. 27(3), 2010, pp. 4-19.
[RAD13]
Radetzki, M.; Feng, C.; Zhao, X.; Jantsch, A. “Methods for Fault Tolerance in Networks
on Chip”. ACM Computing Surveys, vol. 46-1, October 2013, 38p.
[ROD09a] Rodrigo, S.; Medardoni, S.; Flich, J.; Bertozzi, D.; Duato, J. “Efficient implementation of
distributed routing algorithms for NoCs”. Computers & Digital Techniques, vol.3-5,
September 2009, pp.460-475.
[ROD09b] Rodrigo, S.; Hernandez, C.; Flich, J.; Silla, F.; Duato, J.; Medardoni, S.; Bertozzi, A D.;
Mej, A. Dai, D. “Yield-oriented evaluation methodology of network-on-chip routing
implementations”. In: SoC, 2009, pp. 100-105.
[ROD11]
Rodrigo, S.; Flich, J.; Roca, A.; Medardoni, S.; Bertozzi, D.; Camacho, J.; Silla, F.;
Duato, J. “Cost-Efficient On-Chip Routing Implementations for CMP and MPSoC
Systems”. Transactions on Computer-Aided Design of Integrated Circuits and Systems,
vol. 30-4, 2011, pp. 534-547.
[RUS08]
Rusu, C.; Grecu, C.; Anghel, L. “Improving the scalability of checkpoint recovery for
networks-on-chip”. In: ISCAS 2008. pp.2793-2796.
[SAL08]
Salminen, E.; Kulmala, A. ; Hamalainen, T. D. “Survey of Network-on-Chip Proposals”.
White Paper OCP-IP, 2008, 13p.
[SAN10]
Nassif, S. R.; Mehta, N.; Cao, Y. “A resilience roadmap”. In: DATE, 2010, 6p.
[SCH07]
Schonwald, T.; Zimmermann, J.; Bringmann, O.; Rosenstiel, W. “Fully Adaptive FaultTolerant Routing Algorithm for Network-on-Chip Architectures”. In: Euromicro, 2007,
pp. 527-534.
[SEM11]
Sem-Jacobsen, F.; Rodrigo, S.; Skeie, T. “iFDOR: Dynamic Rerouting on-Chip”. In:
International Workshop on Interconnection Network Architecture: On-Chip, Multi-Chip,
2011, pp 11-14.
[SHA14]
Shafik, R. A.; Al-Hashimi, B. M.; Chakrabarty, K. “System-Level Design Methodology.
In Energy-Efficient Fault-Tolerant Systems”. New York: Springer, 2014, pp. 169-210.
[SKE09]
Skeie, T.; Sem-Jacobsen, F.O.; Rodrigo, S.; Flich, J.; Bertozzi, D.; Medardoni, S.
“Flexible DOR routing for virtualization of multicore chips”. In: SOC, 2009, pp. 73-76.
[SOU13]
Sourdis, I.; Strydis, C.; Armato, A.; Bouganis, C.; Falsafi, B.; Gaydadjiev, G.; Isaza, S.;
Malek, A.; Mariani, R.; Pnevmatikatos, D.; Pradhan, D.K.; Rauwerda, G.; Seepers, R.;
Shafik, K.; Sunesen, K.; Theodoropoulos, D.; Tzilis, S.; Vavouras, M. “DeSyRe: Ondemand system reliability”. Microprocessors and Microsystems. 37, 8, 2013, pp 981–
1001.
[STA11]
Stanisavljević, M.; Schmid, A.; Leblebici, Y. “Reliability of Nanoscale Circuits and
Systems - Methodologies and Circuit Architectures”. Springer, 2011, 195p.
90
[SUT13]
Sutter, H. “The Free Lunch Is Over: A Fundamental Turn Toward Concurrency in
Software”.
Available
at:
http://www.gotw.ca/publications/concurrency-ddj.htm.
December 2013.
[TAN06]
Tanenbaum, A. S.; Woodhull, A. S. “Operating systems: design and implementation”.
Pearson Prentice Hall, 2006. 1054 p.
[THO10]
Thonnart, Y.; Vivet, P.; Clermidy, F. “A fully-asynchronous low-power framework for
GALS NoC integration”. In: DATE 2010.
[TIL13]
TILERA.
“TILE-Gx
Processor
Family”.
Available
http://www.tilera.com/products/processors/TILE-Gx_Family. December 2013.
at:
[WAC11a] Wachter, E. W. “Integração de Novos Processadores em Arquiteturas MPSOC: um
Estudo de Caso”. Master Dissertation, Programa de Pós-Graduação em Ciência da
Computação, PUCRS, 2011, 92p.
[WAC11b] Wachter, E. W.; Biazi, A.; Moraes, F.G. “HeMPS-S: A Homogeneous NoC-Based
MPSoCs Framework Prototyped in FPGAs”. In: ReCoSoC, 2011, 8p.
[WAC12a] Wachter, E. W.; Carlo, L.; Carara, E.; Moraes, F.G. “An Open-source Framework for
Heterogeneous MPSoC Generation”. In: SPL, 2012, 4p.
[WAC12b] Wachter, E. W.; Moraes, F.G. “MAZENOC: Novel Approach for Fault-Tolerant NoC
Routing”. In: SOCC, 2012, 6p.
[WAC13a] Wachter, E.W.; Erichsen, A.; Amory, A.M.; Moraes, F.G. “Topology-Agnostic FaultTolerant NoC Routing Method”. In: DATE, 2013, 6p.
[WAC13b] Wachter, E.W.; Amory, A.M.; Moraes, F.G. Fault Recovery Communication Protocol for
NoC-based MPSoCs In: ISVLSI PhD Forum, 2013, 2p.
[WAC14a] Wachter, E.W.; Erichsen, A.; Juracy, L.; Amory, A.M.; Moraes, F.G. “Runtime Fault
Recovery Protocol for NoC-based MPSoCs”. In: ISQED, 2014, 8p.
[WAC14b] Wachter, E.W.; Erichsen, A.; Juracy, L.; Amory, A.M.; Moraes, F.G. “A Fast Runtime
Fault Recovery Approach for NOC-Based MPSOCS for Performance Constrained
Applications”. In: SBCCI, 2014, 7p.
[WAC15]
Wachter, E.; Ventroux, N.; Moraes, F. “A Context Saving Fault Tolerant Approach for a
Shared Memory Many-Core Architecture”. In: ISCAS (International Symposium on
Circuits and Systems), 2015. 4p
[WAN06]
Wang, L-T; Wu, C-W; Wen, X. “VLSI test principles and architectures: design for
testability”. Academic Press, 2006.
[WOS07]
Woszezenki, C. “Alocação de Tarefas e Comunicação entre Tarefas em MPSoCs”.
Dissertação de Mestrado, Programa de Pós-Graduação em Ciência da Computação,
PUCRS, 2007, 121p.
[ZHU06]
Zhu, X.; Qin, W. “Prototyping a Fault-Tolerant Multiprocessor SoC with Run-Time Fault
Recovery”. In: DAC, 2006, pp. 53–56.
91
APPENDIX 1 – LIST OF PUBLICATIONS
Table 13 presents the set of publications held since the beginning of the PhD. The
description column links the paper to this work section, when applicable, or to the main theme of
the publication.
Table 13 – Publications during the PhD period.
Publication
1
2
3
4
5
6
7
8
9
10
11
12
13
HeMPS-S: A Homogeneous NoC-Based MPSoCs Framework
Prototyped in FPGAs
Wächter, E.W.; Biazi, A.; Moraes, F.G.
In: ReCoSoC, 2011.
Exploring Heterogeneous NoC-based MPSoCs: from FPGA to HighLevel Modeling
Ost, L.; Almeida, G.M.; Mandelli, M.; Wächter, E.W.; Varyani, S.; Indrusiak,
L.; Sassatelli, G.; Robert, M.; Moraes, F.G.
In: ReCoSoC, 2011.
Exploring Adaptive Techniques in Heterogeneous MPSoCs based on
Virtualization
Ost, L.; Varyani, S.; Mandelli, M.; Almeida, G.; Indrusiak, L.; Wächter, E.W.;
Moraes, F.G.; Sassatelli, G.
ACM Trans. on Reconfigurable Technology and Systems, v. 5, p. 17:117:11, 2012.
An Open-source Framework for Heterogeneous MPSoC Generation
Wächter, E.W.; Carlo, L.; Carara, E.; Moraes, F.G.
In: SPL, 2012
A Spectrum of MPSoC Models for Design Space Exploration and Its
Use
Petry, C.; Wächter, E.W.; Castilhos, G.; Moraes, F.G.; Calazans, N.
In: RSP, 2012
MAZENOC: Novel Approach for Fault-Tolerant NoC Routing
Wächter, E.W.; Moraes, F.G.
In: SOCC, 2012
Topology-Agnostic Fault-Tolerant NoC Routing Method
Wächter, E.W.; Erichsen, A.; Amory, A.M.; Moraes, F.G.
In: DATE, 2013
Fault Recovery Communication Protocol for NoC-based MPSoCs
Wächter, E.W.; Amory, A.M.; Moraes, F.G.
In: ISVLSI, 2013, PhD Forum, 2013.
Runtime Fault Recovery Protocol for NoC-based MPSoCs
Wächter, E.W.; Erichsen, A.; Juracy, L.; Amory, A.M.; Moraes, F.G.
In: ISQED, 2014.
A Framework for MPSoC Generation and Distributed Applications
Evaluation
Castilhos, G.; Wächter, E.W.; Madalozzo, G.A.; Erichsen, A.; Monteiro, T.;
Moraes, F.G.
In: ISQED, 2014.
A Fast Runtime Fault Recovery Approach for NOC-Based MPSOCS for
Performance Constrained Applications
Wächter, E. W.; Erichsen, A.; Juracy, L.; Amory, A. M; Moraes, F. G.
In: SBCCI, 2014.
A Context Saving Fault Tolerant Approach for a Shared Memory ManyCore Architecture
Wachter, E.; Ventroux, N.; Moraes, F.
In: ISCAS, 2015. 4p.
An Integrated Method for Implementing Online Fault Detection in NoCBased MPSoCs
Fochi, F.; Wächter, E.; Erichsen, A.; Amory, A.; Moraes, F.
In: ISCAS, 2015. 4p.
Description
Prototypingofthe
referencearchitecture
Analysisofdifferent
applicationsin
differentorganizations
ofheterogeneous
MPSoCs
Analysisofdifferent
applicationsin
differentorganizations
ofheterogeneous
MPSoCs
Frameworkofthe
referencearchitecture
Analysisofthe
differentmodelsofthe
referenceMPSoC
Chapter4
Chapter4
FirstMethodof
Chapter5
FirstMethodof
Chapter5
Toolsettogeneratethe
referencearchitecture
SecondMethodof
Chapter5
Approachdevelopedin
thesandwichthesis–
Chapter6
Exampleofpossible
faultdetectionto
exploreinChapter3